BME Masters Theses 2017-2018

3/30/2017 BME
Master's Theses 20172018
Departement Werktuigkunde > BME Master's Theses 20172018 > Proposals
Proposals
Table of Contents
Biomechanics
Biomedical Data Processing
Biomedical Imaging
Biomedical Instrumentation
Biomaterials
Tissue Engineering
1. Biomechanics
Mechanical characterization of soft biological tissue: Biaxial testing with digital image correlation –
#1
by Fehervary Heleen – last modified Mar 30, 2017 11:38
Discipline: Biomechanics
Promotors: Jos Vander Sloten , Nele Famaey – Mentor: Heleen Fehervary
Number of students: 1
Program of study option: Any
Background
A better understanding of the biomechanical behavior of biological tissue is essential for society, as it provides the medical community with the
engineering advice needed to enhance safety and quality in patient treatment. The performance of biomedical devices or procedures is increasingly
assessed by means of finite element simulations. Using the correct material properties is essential for the outcome of the simulation: for example when
choosing the correct size for a stent, not only the patient’s geometry but also the material properties of the artery influence the result.
Material properties are characterized using experimental setups, such as a planar biaxial tensile test or an extensioninflation test. During an experiment
loads and deformations are tracked which are then used for material modelling and parameter fitting. Digital Image Correlation (DIC) is a stateoftheart
technique to track the deformation field of the tissue using two camera’s and a speckle pattern.
Currently the Biomechanics Section (BMe) is opening the ‘Leuven Biomechanical Testing Facility’ which possesses, along with different mechanical set
ups, a DIC system that will allow improved mechanical testing and also creates new opportunities such as inverse finite element parameter fitting.
Some initial experiments and measurements with this DIC system have been performed, but the challenge of this thesis is to develop and optimize a
complete protocol to use this DIC system with the planar biaxial setup. This requires the setup of the camera’s, the application of a speckle pattern,
the processing of the images and the implementation of the deformation measurements into a parameter fitting code.
Objectives
The purpose of this thesis is to develop and optimize a dedicated protocol for the use of a DIC system with a planar biaxial test.
This thesis will involve:
Creation of a speckle pattern onto biological tissue: initial experiments have been carried out, but different materials and techniques should be
explored to optimize the application onto the tissue and the size and frequency of the speckles.
Image processing: using the dedicated software of the system to investigate the different effects of parameters and to provide a suitable output.
Implementation of the results from image processing into a parameter fitting (PF) code:
Conventional PF: this requires homogenizing the deformation measurements. The PF code is available for this thesis (developed in Matlab).
Inverse PF: the experiment is simulated using the Finite Elements (FE) software Abaqus, in which the deformation field is applied to the sample. The
challenge is not the simulation itself, but the development of an automatic procedure. Abaqus offers a userfriendly scripting environment to this end.
Inhomogeneity from Experiment Corrected (IEC) PF: a new parameter fitting procedure developed by BMe, that is a combination of the two above.
Currently a master thesis student is implementing this procedure. The task here is to include the possibility of using a deformation field.
Finally the performance of the three parameter fitting codes should be compared in results, computation time, accuracy, …
This thesis gives the student the opportunity to work in three different domains: practical work, image processing techniques, programming and finite
element modelling.
Additional Info:
https://www.mech.kuleuven.be/mpbme1718/proposals/fulltext 1/52
3/30/2017 BME Master's Theses 20172018
Since the student will be working with biological tissue, an HSE check is required.
Complete biomechanical characterization of human thoracic aorta – #11
by Vastmans Julie – last modified Mar 30, 2017 11:25
Promotors: Nele Famaey , Jos Vander Sloten – Mentors: Julie Vastmans , Emma Vanderveken (KU Leuven)
Background
Cardiovascular diseases (CVDs) are the leading cause of death globally. Examples of this type of disease are thoracic and abdominal aneurysms,
atherosclerosis, etc. Treatment of CVDs consists either of drug treatment or surgical intervention. Evaluating new surgical techniques and medical
devices, e.g. stents, require extensive in vivo and in vitro testing. Numerical models, like finite element (FE) modeling, provide a useful tool to aid the
development of medical devices, and even to provide patientspecific solutions. However, these numerical models require knowledge about material
properties (both mechanical and microstructural properties) of the modeled tissues. Both healthy and diseased aortic tissues have already been
investigated. For example, thoracic aortic aneurysms have been mechanically tested at the Biomechanics Section. However, no studies could be found
that investigate the properties of the thoracic aorta both mechanically and microstructurally along its complete length, from aortic valve to diaphragm.
Currently, new stateoftheart testing equipment is being installed in the Biomechanics Section (BMe). Aditionally, 10 complete human thoracic aortas,
ranging from the aortic valve to the diaphragm, are being collected at the Vesalius institute. These aortas are dedicated for mechanical testing combined
with histological examination. The availablility of this kind of tissue and the proper testing facilities is highly unique, and expected to generate a wealth
of information.
Objectives
The objective of this thesis is to fully characterize the human thoracic aorta.
This thesis will invove
Mechanical experiments performed on the new planar biaxial setup. These mechanical experiments will be performed concurrently with
histological examination. (The histological examination will be done at UZ Leuven, and is not part of this thesis)
Determining the material properties based on the performed mechanical experiments, using nonlinear parameter optimization schemes (for which
Matlab code is available).
Development of a FE model of the human thoracic aorta, with the correct material properties
This thesis takes place in the Soft Tissue Biomechanics group (https://www.mech.kuleuven.be/en/bme/research/softtissuebiomechanics) and allows
the student to work in an interdisciplinary environment: engineers, biomedical scientists and surgeons are closely involved in this project. Finally,
considering the uniqueness of the collected data, it is expected that the results obtained in this thesis will result in a conference paper or journal
publication.
Additional Info:
This project involves the handling of human tissue samples. It is, therefore, compulsory that the researchers have the required vaccinations uptodate,
which should be doublechecked with idewe. If you have any questions in this respect, please contact the supervisors.
Comparison of a commercial and patientspecific clavicle fracture fixation plate: a computational
study – #10
by Vancleef Sanne – last modified Mar 24, 2017 16:34
Promotors: Jos Vander Sloten , Ilse Jonkers – Mentor: Sanne Vancleef
Number of students: 1 or 2
Background
Operative fixation of clavicle fractures has a high complication rate. It has been hypothesized that the complication rate is related to the fixation plate
design. Nowadays, different clavicle fracture fixation plate designs are available, but they are far from ideal. Therefore, a patientspecific fixation plate
was designed.
Finite Element (FE) models are widely used in orthopedic implant design. Advantages of this technique are that 'what if' scenarios can be easily
simulated. However, to properly evaluate the stress and strain distribution of different implant designs, more physiological finite element models are
necessary. In order to create these, musculoskeletal models are often used to asses muscle and joint reaction forces in different activities.
Goal
The goal of this study is to evaluate the patientspecific fixation plate during activities of daily living by comparing it to commercially available fixation
plates. In order to do this, a more physiological finite element model should be created, which includes patient specific material properties and
physiological boundary and loading conditions.
Clinical relevance
Clavicle fractures are common fractures, they account for 4% of all fractures and 44% of all shoulder injuries. There is still discussion whether to apply
conservative or operative treatment, since current fixation plates are relatively thick and can cause soft tissue irritation. Furthermore, since the clavicle
is a subcutaneous structure, thick implants can cause irritation when wearing a backpack and often a second operation is needed to remove the
implant. Therefore, thin plate implants would be beneficial to reduce these side effects.
Work description
Calculating muscle and joint reaction forces for different activities of daily living using a musculoskeletal model.
Create a Finite Element model in Abaqus of a clavicle fracture with patientspecific material properties, physiological boundary and loading conditions.
Evaluate stress and strain for different fixation plate designs.
Optimization of the patientspecific implant, by using topology optimization.
Predicting the growth and remodeling of arteries – #12
by Vastmans Julie – last modified Mar 30, 2017 10:09
Promotors: Nele Famaey , Jos Vander Sloten – Mentors: Julie Vastmans , Lauranne Maes (KU Leuven)
Background
Biological soft tissues, such as arteries, are alive! This poses extra challenges, which do not exist in classical engineering materials. The most
prominent of these challenges is the capability of arteries to grow and remodel. These growth and remodeling processes take place during the patient’s
entire lifetime. These processes make that, even when unpressurized, an artery is in a complex loaded state, often referred to as the ‘residual stress
state’.
Furthermore, growth and remodeling processes take place in the event of a changing mechanical environment of the artery. The mechanical
environment of an artery can change during a surgical procedure, such as stenting. Another example of such a surgical procedure is the Ross
procedure, where surgeons place the pulmonary valve in aortic location, subjecting the pulmonary autograft to higher pressures than normal.
A simplified version of this procedure is currently being performed on sheep, after which the different tissues (pulmonary artery and aorta) are
mechanically tested. The preliminary mechanical tests showed that the pulmonary artery starts to behave like an aorta when placed in aortic position,
showing that growth and remodeling processes took place.
In the previous two years, the soft tissue biomechanics group developed both an algorithm to calculate the residual stresses, as well as an
implementation of a growth and remodeling model.
Objectives
The objective of this thesis is to further develop a parameter fitting framework for the growth and remodeling algorithm, and to combine it with the
developed residual stresses algorithm. Also, the developed framework will be tested on data obtained from a patient undergoing a procedure during
which the patient’s dilated aorta was wrapped with an external support.
This thesis will involve
Development of a parameter fitting framework of a growth and remodeling algorithm combining Matlab, Python and Abaqus (finite element
software)
Processing of preoperative and postoperative images of a patient in Materialise software (Mimics) and creating a patientspecific finite element
model
During this thesis, the student will gain expertise on the important domain of nonlinear finite elements and continuum mechanics and on common
engineering programming languages. Due to the challenging nature of this thesis, the student will be strongly supported by the soft tissue biomechanics
group (https://www.mech.kuleuven.be/en/bme/research/softtissuebiomechanics). The developments will be directly useful to many of the research
topics currently ongoing in our group, and is therefore likely to lead to at least one journal publication.
Nanoscale force sensors to measure cellular forces – #13
by Vaeyens MarieMo – last modified Mar 30, 2017 10:30
Promotor: Hans Van Oosterwyck – Mentor: MarieMo Vaeyens
Background:
Forces are everywhere, at every scale, playing their role in the bigger part. The molecular forces at the interface between cells and their micro‐environment in the
body are influencing processes not only at molecular and cellular scale but also at larger scales (��ssue, organ, body). At the cellular level, these forces affect cell
adhesion, migra��on, prolifera��on and differen��a��on, and on a larger scale they affect processes such as blood vessel forma��on and cancer metastasis. Because
cell‐matrix mechanics is crucial for physiological and pathophysiological processes, these forces are being studied in our group with the ul��mate goal of developing
novel strategies in ��ssue engineering and cancer treatment.
Endothelial cells that form new blood vessels probe the mechanical proper��es of their environment ‐ the extracellular matrix (ECM) ‐ by applying trac��onal forces.
Larger trac��on forces are o擌�en located in focal adhesions where cellular adhesion molecules such as vinculin are bundled, clutching the cell to the substrate and
allowing force transmission from the cytoskeleton to the hydrogel. At these focal adhesions, mechanical signals are transduced and converted into a biochemical
response.
Some current high technological approaches to measure molecular tension across cell‐surface or intracellular molecules are based on Förster resonance energy
transfer (FRET). FRET‐force sensors allow es��ma��ng forces at the protein level by for example introducing quenched fluophores and a calibrated nanospring part
into protein. When the nanospring is stretched upon applica��on of force, the fluorophores are unquenched, resul��ng in emission of a reduced fluorescence signal
that can be related to the force needed to stretch the nanospring.

Content:
A biosensor will be used to study vinculin dynamics and forces at focal adhesions of endothelial cells, in func��on of mechanical s��ffness and in func��on of cellular
force inhibitors. In a first phase the FRET biosensor will be validated by measuring live cells migra��ng on a range of substrates with defined mechanical proper��es.
To carry on this project, poly acrylamide (PAM) hydrogels will be used as substrate. PAM has been extensively used as a substrate surrogate due to its transparency,
tunable s��ffness and ease of handling. The fluorescence emission of the sensor will be measured with confocal fluorescence microscopy and compared to results
from trac��on force microscopy (TFM) analysis from beads in the sample. In a second phase the validated sensor will be used to map the vinculin forces while
chemically disrup��ng the cell’s force genera��ng capabili��es. This mul��disciplinary thesis subject aims at combining biosensor and hydrogel engineering, cell culture
experiments, cell mechanics, fluorescence microscopy, op��cs, image analysis with matlab, cell biology, ��ssue engineering and biomaterials, within the broader
scope of ul��mately regula��ng angiogenesis in novel ��ssue engineering strategies and cancer treatment. Depending on the student’s interest, the focus of the thesis
can be adapted.
Fig. 1 A vinculin FRET force sensor expressed by an endothelial cell. (A) Confocal fluorescence microscopy image of the endothelial cell showing vinculin (red) and F‐
ac��n (grey). (B) The FRET‐efficiency of the vinculin FRET force sensor. (by Marie‐Mo Vaeyens)
More info...
Identification of physical muscle weakness – #17
by Tanghe Kevin – last modified Mar 28, 2017 08:51
Promotor: Friedl De Groote – Mentor: Kevin Tanghe
Background
In the developed countries, the age of the population is continuously increasing. More and more people are becoming less mobile and need assistance
for daily activities. Robotic devices can be used to provide this assistance.
In the MIRAD project (http://www.miradsbo.be/), an exoskeleton for the lower limbs is under development. The envisioned patient groups are people
suffering from muscle weakness, cerebral palsy, multiple sclerose … The patients are not paralyzed, they can still move their legs.
An important task in delivering support is determining the assistance to be supplied to the subject. If too little assistance is given, the subject may not
be able to execute a desired movement. If too much assistance is given, there is a risk of further muscle weakening in the long term. The MIRAD
project thus aims to provide assistanceasneeded (AAN). The computation of this AAN is based on a model of the weakness of the patient. The
identification of this patientspecific model is one of the goals of this thesis.
Goal
The goal of this thesis is twofold. A first goal is to identify a musculoskeletal model of a specific patient with muscle weakness. The force that a muscle
can exert has a known relation to the position of the limbs. Therefore, it would be beneficial to model the weakness at the muscle level (i.e. what is the
maximal force that each muscle can deliver?). In this case, the weakness can be computed for every position of the limbs. Another possibility is to
define weakness at joint level (i.e. what is the maximal torque that can be exerted at the joint?), but then the relation to the kinematics is lost. This
would imply that the parameters for the weakness should be different for different positions of the limbs.
The second and equally important goal is to determine how the identification of weakness can be done with a practical clinical experiment, i.e. without
the need of extensive measurements or very specialized equipment.
Work description
In this thesis the student will need to
Do a literature study on musculoskeletal modeling and clinical test
Define which parameters will be identified for the patientspecific model
Define a set of clinical tests which allow to do the identification
Do experiments: perform the clinical tests and identify the patientspecific model
Validate the results
Kinematic reconstruction of cervical spine motion – #38
by Vander Sloten Jos – last modified Mar 28, 2017 18:13
Promotors: Jos Vander Sloten , Jan Goffin (UZ Leuven, neurosurgery) , Lennart Scheys – Mentors: Hannelore Boey , Tiago DE MELO LOPES
MARTINHO MALAQUIAS
The Neurosurgery Department has a unique database of fluoroscopic data of human volunteers, performing neck flexionextension,
lateroflexion and axial rotation. Data are available of appr. 80 volunteers, of different age categories.

The aim of this master thesis is to perform a kinematic analysis of the cervical spine segments, identifying translation and rotation
patterns in three dimensions from the fluoroscopic image data. This should improve our understanding of the impact of the degeneration
process on motion patterns, subsequent changes in mechanical loads on motion segments and what criteria an ideal disc prosthesis should
fulfill to restore physiological motion. Motion analysis software is available from a previous PhD thesis (Joris Walraevens). The student
will expand this software to improve the degree of automation in order to process the large database. Next, three dimensional kinematic
motion patterns have to be derived and presented in a format that allows for easy interpretation in collaboration with the neurosurgeons.
The master thesis will run in collaboration with the Institute for Orthopaedic Research and Training.
Stretching the limits of soft biological tissue testing – #37
by Kapeliotis Markos – last modified Mar 30, 2017 12:03
Promotors: Jos Vander Sloten , Nele Famaey – Mentors: Markos Kapeliotis , Heleen Fehervary
Legal Notice: The student(s) to whom this master thesis subject is assigned, will have to renounce intellectual property rights
Background
Soft biological tissues are nonlinear hyperelastic and anisotropic. In many situations in biomechanics, we want to simulate how these tissues respond
to mechanical loading. Research areas such as accident damage prevention use finite element models of the head to simulate the effect of an impact.
Using the correct material properties of the tissues in this head is clearly essential for the outcome of the head impact simulations. Hence, accurate
mechanical testing of these tissues is essential.
Bridging veins (BVs) drain blood from the cerebral cortex into the sinus sagittalis superior and by doing so they form a ‘bridge’ over the subdural space.
Acute subdural hematoma is a head lesion with high mortality rate and one third of the cases of ASDH is caused by a rupture of bridging veins. To
improve the head models that are used to predict head trauma, the mechanical properties of these bridging veins must me known. However, classical
test setups are not suitable for these tissues due to their small dimensions.
Therefore, these properties are characterized using a novel experimental setup, a biaxial tensile test bench especially designed and fabricated in our lab
for small sized tissues. However, due to the boundary conditions of the experimental setup, inaccuracies are introduced in the experimental data.
Simple parameter fitting does not take these inaccuracies into account, which would make the resulting material properties unreliable. Inverse finite
element (FE) parameter fitting does take these inaccuracies into account; however, this method is computationally very expensive.
At the Biomechanics section (BMe) a new method for parameter fitting was developed (Fehervary et al., 2016), which combines the best of the two
approaches mentioned above: it accounts for the inaccuracies, but is not as computationally expensive as traditional inverse FE. This method was
tested on a classical biaxial setup, but not for the novel setup described above.
Objectives
The purpose of this thesis is to construct a framework for the use of the new parameter fitting method with novel biaxial setup for small tissue samples.
This framework will make use of Matlab, python and Abaqus.
Developing FE models that simulate the experimental conditions
Image processing of the experimental data such as edge detection and digital image correlation. Initial Matlab code has been developed at BMe.
Implementation of the results from image processing in Abaqus. The challenge is not the simulation itself, but the development of an automatic
procedure. Abaqus offers a userfriendly pythonbased scripting environment to this end.
Post processing of the simulated and experimental data in Matlab, which will involve the comparison of the data and the formulation of a stop
criterion.
Application of the developed framework to the available experimental data and evaluation of the new method.
This project will give you the opportunity to gather a wide range of theoretical and practical skills, including software manipulation and handson
experimental work. The results of this thesis are likely to lead to a publication.
Biomechanical assessment of thoracic ascending aortic aneurysms – #34
by Farotto Dario – last modified Mar 29, 2017 17:24
Promotors: Nele Famaey , Jos Vander Sloten – Mentor: Dario Farotto
Background
Aneurysm is a serious and common disease which consists of a localized dilatation of the aorta resulting from compromised structural integrity. The
primary treatment available is a highlyinvasive elective surgery. The current criteria for surgery are based on size and perform poorly. Wall stress
estimation through biomechanical simulations represents a promising tool to provide a more individual risk assessment and have the potentiality to
reduce not only mortality but also unnecessary surgeries.
Objectives
The current project is going to be developed in this framework and will build up from the previous work of Master students and PhD’s. The purpose is to
design a method to estimate the material properties of the aorta in a noninvasive way. Material properties are fundamental for accurate results of
mechanical simulations.
For this study, data has already been collected from patients that were scheduled for surgery for ascending aortic thoracic aneurysm. The data
comprehend 4D CT data of the aorta before the operation and samples of the resected tissues. The material properties of the resected tissues are
estimated with planar biaxial test to validate the noninvasive method.
This thesis will therefore involve:
Processing and analysis of the preoperative images to obtain the patient specific geometry with Materialise software (Mimics Innovation Suite).
Estimation of mechanical properties of the aneurysm based on noninvasive measurements (based on python)
Finite element simulations of the aneurysms with the patientspecific geometry and the parameters obtained from biaxial tests and from noninvasive
measurements (Abaqus).
This thesis gives the opportunity to learn and deepen the student’s knowledge about Python and finite element simulation (Abaqus). This thesis takes
place in the Soft Tissue Biomechanics group (https://www.mech.kuleuven.be/en/bme/research/softtissuebiomechanics) and allows the student to work
in an interdisciplinary environment: engineers, biomedical scientists and surgeons are closely involved in this project. Objectives can be adapted on the
interests of the students. The results of the project could lead to the publication of a paper.
Biomechanical characterization of the mitral valve – #51
by Famaey Nele – last modified Mar 30, 2017 09:59
Promotors: Jos Vander Sloten , Nele Famaey – Mentor: Patricia Silvana Da Torre Lopes (KU Leuven Materialise)
Background
Mitral valve disease is the most commonly occurring valvular abnormality in the western world and its incidence is expected to rise as life expectancy
increases worldwide.
Mitral regurgitation arises from inadequate coaptation of the valve leaflets during ventricular compression, resulting in backflow of blood to the left
atrium. Mechanisms for mitral regurgitation can be subdivided into organic, characterized by structural abnormalities of the leaflets of subvalvular
apparatus, and functional, often associated to heart failure and the dilation of the left ventricle that pulls the leaflets apart.
While the standard treatment for mitral regurgitation involves openheart surgery, up to 50% of the patients are denied surgery. Transcatheter mitral valve
devices are an alternative solution for patients who are too frail to undergo surgery. However, several challenges are associated to the design of such
devices. To guarantee their effectiveness and safety, a solid understanding of the mitral valve anatomy, dynamics and biomechanics is, therefore,
crucial.
Objectives
The goal of this project is the in vitro assessment of the biomechanical properties of degenerated and healthy human mitral valve. Clinical 4D CT and
4D echocardiography are available for a pilot set of patients and excised tissue samples are being collected.
Your contribution will include:
the analysis and processing of medical images (4D CT and 4D echo)
biaxial mechanical testing and histological analysis of resected human mitral valve tissue
implementation of these findings into a realistic finite element model
This information will be incorporated into a statistical shape model of the mitral valve. These computational models are being developed in a
collaborative project carried out at Materialise, the KU Leuven and the University of Antwerp, and are aimed at the design optimization of transcatheter
mitral valve devices.
This project will give you the opportunity to gather a wide range of theoretical and practical skills, including software manipulation and handson
experimental work.
Additional Info:
3D Force Microscopy of Cells – #52
by Jorge Peñas Alvaro – last modified Mar 30, 2017 10:11
Promotor: Hans Van Oosterwyck – Mentor: Alvaro Jorge Peñas
Background:
The importance of cell biomechanics
The cell is a complex and dynamic entity, which continually adapts and responds to stimuli from its environment. Traditionally, research efforts have
been focused on understanding cellular response to chemical signals. However, increasing interest in cell mechanics has been recently generated, as it
has been shown to regulate cell behavior. Indeed, biomechanics plays a key role in a number of fundamental physiological and pathological processes
of living organisms. For example, endothelial cells of inner arterial walls feel the shear stress caused by the blood flow and promote the remodeling of
the artery diameter adapting it to the flow. Furthermore, it has been shown certain correlation between changes in the mechanical properties and some
widespread diseases such as cancer, malaria, anemia, asthma, osteoporosis, or heart failure. The previous examples show that cells are capable of
sensing and responding to the mechanical cues of their surrounding environment converting them into biochemical signals that, ultimately, affects many
biological processes, including cell migration and growth. This physical interaction between cells and their surrounding environment is driven by cellular
forces; thus, the quantification of these forces provides insight about normal and diseased tissue.
Traction Force Microscopy
Traction Force Microscopy (TFM) is a multidisciplinary technique that combines optical microscopy, image processing and mechanics to estimate
cellular traction forces. TFM relies on measuring the deformations that cells exert on their surrounding environment and use them to recover cellular
tractions by solving an inverse problem. In other words, the idea behind TFM is to find the causes (cellular tractions) from the observed effects
(deformations of their environment). Examples of these types of inverse problems can be found in many fields of science, including astronomy, medicine
(Xray, computerized tomography), image analysis (deconvolution) or economics (financial mathematics).
Towards reliable 3D TFM
TFM was originally developed for 2D experiments, where cells are cultured on the surface of synthetic polyacrylamide hydrogels. These types of gels
present a wellknown mechanical behavior that has been used in combination with the cellinduced deformations to recover cellular tractions in a wide
range of research studies. However, cells tend to lose many of their normal functions when cultured in 2D. Thus, in order to better approximate to in
vivo conditions, efforts are being made towards a reliable recovery of 3D tractions exerted by cells embedded in 3D complex environments, whose
mechanical characterization and model are missing.
Content (objective and work description):
The goal of this project is to establish and test a methodology to quantify the 3D traction fields exerted by cells inside complex, realistic, nonlinear
environments, such as (fibrillar) collagen gels. The work will combine lab experiments with computer –based analyses. The student will prepare cell
compatible collagen gels (protocols available) and mechanically characterize them by a set of techniques, including uniaxial stretching and
shear/extensional rheometry. Obtained information will be fed to a recently developed software (semiaffine elastic network optimizer, SAENO) that
makes use of a customized mechanical model of collagen gels to recover 3D cellular tractions. In this context, the student will be expected to test
SAENO software and include it in a general TFM pipeline that will be used to study cellular mechanics under different stimuli.
Depending on the student’s interest, the focus of the thesis can be adapted.
Figure Caption:
Local deformations of collagen gels around cells are used recover cellular tractions. Displacement directions (arrows) of collagen fibers around an
endothelial cell (white) for a selected XY plane (A); computed 3D full field displacements (in µm) induced by the cell (B).
More info...
Quantification of bone and cartilage structure in the human wrist – #55
by Mys Karen – last modified Mar 30, 2017 13:56
Promotor: Harry van Lenthe – Mentor: Karen Mys
Bone diseases (e.g. osteoporosis, osteoarthritis) and trauma (e.g. fracture) have a highsocio economic impact. Due to the rising life expectancy, the
socioeconomic impact will increase further. All these diseases influence the bone microarchitecture as well as the bone mechanical properties.
The CMC1 is one the joints in the wrist (between metacarpal 1 and trapezium) where osteoarthritis develops frequently (Fig 1); yet, a detailed
understanding of why osteoarthritis is developing at this joint is lacking. It is believed that mechanical loading plays a role, and that this is being
reflected in the structure of the joint. In a preliminary study we have demonstrated (1) that there is a correlation between cortex thickness and cartilage
thickness, and (2) that due to osteoarthritis not only the shape of the trapezium is changing (as already mentioned in literature), but the shape of
metacarpal1 is changing too. The preliminary study used 3D in vivo CBCTimages of the whole wrist and microCTimages of the trapezium bone. In
order to get a better estimation of the cartilage thickness over the entire bone surface, a 4D CBCTscan is needed.
Aim 1: Determine the cartilage thickness of the CMC 1joint using 4D CBCT.
Determine the cartilage thickness of the trapezium bone based on microCTimages
Beside the cartilage thickness, determination of the cortex thickness is relevant. Software has been developed in our group that can calculate
automatically a mask of the cortical bone and a mask of the trabecular bone based on microCTimages.
Aim 2: Determine the cortex thickness using CBCT. This will require to expand our current tools to automatically identify the cortical bone in CBCT
images.
Spatial correlation of cortex thickness to cartilage thickness.
Aim 3: Calculate the shape of the metacarpal and the trapezium bone, microstructural parameters,… and correlate this with cortex thickness and
cartilage thickness.
Additional Info:
Working with human tissue
Validation of boundary conditions to create a highresolution FEmodel of a trapezium bone – #54
Bone diseases (e.g. osteoporosis, osteoarthritis) and trauma (e.g. fracture) have a highsocio economic impact. Due to the rising life expectancy, the
socioeconomic impact will increase further. All these diseases influence the bone microarchitecture as well as the bone mechanical properties.
The CMC1 (joint between metacarpal 1 and trapezium) and the STT (joint between trapezium and scaphoid) are joints in the wrist where osteoarthritis
develops frequently (Fig 1); yet, a detailed understanding of this process is lacking. Therefore, an improved knowledge of the mechanical properties on
the microstructural level is desired.
MicroCTbased finite element (FE) analyses have shown to provide an accurate assessment of the mechanical properties of trabecular bone structures.
To be able to develop realistic FEanalyses, realistic boundary conditions need to be defined.
The aims of the thesis are
(1) to define patientspecific realistic boundary conditions for FEmodels based on in vivo CBCTscans
(2) to validate the boundary conditions by comparing the FEmodels with experiments (Fig 2).
This thesis is 50% experimental and 50% computational and gives the student the opportunity to work in close collaboration between engineers and an
orthopaedic surgeon.
Additional Info:
Working with human tissue
Accident Reconstruction to increase the biofidelity of bicycle helmet testing standards. – #58
by Kapeliotis Markos – last modified Mar 30, 2017 12:12
Promotors: Jos Vander Sloten , Guido de Bruyne (Lazer Sport) – Mentors: Dimitris Zouzias (Ku Leuven / Lazer Sport) , Michel Woering ,

Markos Kapeliotis
Current helmet testing standards take only linear acceleration of the brain into account. At this moment, different international research programs aim at
integrating research findings on brain injury mechanisms that are caused by oblique impact. These integrated research findings may lead to new helmet
standard testing procedures. Unfortunately, it remains difficult to predict brain injuries that are seen in real life accidents. Accident reconstruction may
help to find this insight.
Accident reconstruction uses rigid body models or Finite Element (FE) models to reconstruct bicycle accidents, based on a databases of cycling
accident descriptions and medical information. MADYMO (MAthematical DYnamic Models) is used for the rigid body simulations and is a stateofthe
art engineering tool for calculating crash kinematics, which can be coupled with medical information to establish injury thresholds. These thresholds can
be used to improve helmet test standards and thus helmet safety.
Finally, in order to get more insight into what happens inside the head a stateoftheart FE head model will be used. The acceleration profiles obtained
from helmet testing and accident reconstructions will be used as an input. The outcome from these simulations could provide valuable insight for the
testing process and more sophisticated helmet design.
This project allows you to work with FE and MADYMO, while getting experience with a fast mathematical simulation technique that is often used by car
manufacturers. Additionally, a stateoftheart drop test installation of Lazer Sport will be used.
Objective:
The goal of the thesis is to investigate bicycle accidents for understanding fall kinematics and head impact.
Deliverables
Real life accident reconstruction with MADYMO
Input variables for an improved helmet testing process
Correlation matrix between injuries’ predicted with experimental acceleration profiles of the proposed testing standard and acceleration profiles
predicted with MADYMO.
Getting a cell to walk from the inside: Computational modeling of how chemical regulation affects
cell migration – #59
by Vargas Arango Diego – last modified Mar 30, 2017 11:07
Promotor: Hans Van Oosterwyck – Mentor: Diego Vargas Arango
Key terms: cell migration, computational modeling, cell adhesion, mechanotransduction, cytoskeletal dynamics, cell signaling
Background:
Different cell types have been found to migrate under different conditions and stimuli, usually confined to morphogenesis and ending with differentiation
but also in healing and disease. There are multiple modes of cell migration, but one of the most well understood one is single cell mesenchymal
migration. This motility mode is characterized by a cyclic motion consisting of cell polarization, in which the cell forms a (a) protrusive front, then
forms new (b) adhesions to the substrate, and (c) retracts its rear (Figure 1A).
The protrusive front consists of extension of the cell membrane through the polymerization of actin. Attachments or focal adhesions consists of protein
complexes that link the internal cell cytoskeleton with the outside of the cell; these complexes are responsive to physical forces. Finally, retraction
occurs via contraction of internal actin fibers (i.e. stress fibers) via crosslinking myosin molecular motors. Cells are highly active systems that interact
with their environment, and all processes listed (a, b, and c) are linked via chemical signaling and mechanotransduction (the conversion of mechanical
forces into chemical signals). Figure 1B shows the distribution of some of the molecules involved in the migration process.
The objective of this project is to fine tune a multiscale computational model that incorporates the effect of signaling and mechanotransduction into cell
migration. Within the scope of the lab we later hope to model multiple cells and collective migration. Collective migration is essential in building complex
tissues, as in the process of formation of blood vessels from existing ones, a process known as angiogenesis (the lab’s favorite). The understanding
this will bring can help us better develop tissue engineering constructs for vascular healing. Understanding collective motion can also lead to better
strategies to inhibit it during cancer metastasis.
Content:
We have created a multiscale model of the cell using discrete element method (DEM), a method that numerically computes the motion of a large
number of small particles. The particles are used to make a mesh describing the cell surface of particles connected via elastic elements (1). This
triangulated flexible surface allows local (spatial) variations in surface concentration of proteins and mechanical properties of the cell (Figure 1C). The
particles also allow us to discretize interactions between the cell and its surrounding, providing a way to model discrete focal adhesions.
For each particle in the mesh there is an integrin concentration which is allowed to either diffuse along the cell membrane or form focal adhesions when
in contact with the substrate. Stress fibers connect focal adhesions in different parts of the cell. A set of parameters that the student will have to vary
describe the dynamics with which focal adhesions are formed and how attachment force affects lifetime of the adhesions.
With the model we want to address questions of the following sort: How does a force feedback at focal adhesions affect cell migration? What
molecules in the regulatory signaling are most crucial to a cell maintaining a preferred direction of motion? Etc.
Although the model is implemented in Python, and a parameter study will have to be performed, little work should go into changing the model. Initially
the student should only vary parameter values and run simulations. The output, consisting of position of each particle constituting the cellular mesh and
corresponding molecular concentrations, can be analyzed in the language of the student’s choice, for exampleMATLAB, Python, C/C++, etc. The
student will measure the impact of changing the parameters on the speed and direction of cellular motion and compare it to the values obtained for real
cells both in our lab and in the literature (Figure 1D).
However, based on the student’s interests and progress, the model can also be expanded or modified, for example including different cellular pathways
describing the interaction of the different regulatory mechanisms (a, b, and c) in the form of ordinary differential equations (ODEs). There would be
ample space for the student to explore scenarios and conditions that pique his/her interest.
References:
1. Odenthal T, Smeets B, Van Liedekerke P, Tijskens E, Van Oosterwyck H, Ramon H. Analysis of initial cell spreading using mechanistic contact
formulations for a deformable cell model. PLoS Comput Biol. 2013;9(10):e1003267.
2. Reig G, Pulgar E, Concha ML. Cell migration: from tissue culture to embryos. Development. 2014 May 15;141(10):1999–2013.
Figure caption:
Figure 1. (A) Current view of the cell migration cycle. The following processes are highlighted through coloring: actin polymerizationdependent
protrusion (red), cellsubstrate adhesions (purple), and retraction (green). (2) (B) Schematic of the molecules involved in pathways regulating migration. It
shows the different spatial localization in a polarized cell. (2) (C) Rendition of simulated cell. LEFT Particles used to define the cell are shown in white
(i.e. triangle vertices). The mesh is colored according to integrin (X4_n) concentration [molecules/node] extrapolated to the triangles based on area.
RIGHT Bottom view of the cell showing the interface between cell and substrate. (D) Two different scenarios. In the top a cell's migration is not very
directed due to high Rac concentration. Below the opposite case. (2)
More info...
Design and development of a mockup of the human knee joint – #66
by Labey Luc – last modified Mar 30, 2017 13:34
Promotor: Luc Labey – Mentor: Luc Labey
Background:
The human knee is a complex joint. It’s biomechanics is governed by the interaction between three bones (femur, tibia and patella) linked together by
multiple ligaments and other soft tissue structures (such as the menisci). As a consequence, there are currently no validated physical models of this
joint available with similar kinematics and kinetics as a real, native knee joint.
This hinders progress in the treatment of knee pathologies (such as osteoarthritis) and injuries (such ACL tears). Indeed, orthopaedic surgeons currently
always need human cadaver specimens to test new surgical techniques or implants before using them on patients. However, the availability of such
specimens is quite limited, working on them is not very pleasant and evaluating their performance is not easy, as they are not equipped with sensors to
measure changes in kinematics, strains, pressure, …
Objective:
The availability of a biomechanically equivalent mockup of a human knee joint would be a great step forward in this respect. Such a device could be
used by surgeons as well as engineers to analyse new surgical techniques or new implant designs and see how they change the natural kinematics and
kinetics of a knee joint (at least in a first phase). It may also be useful for surgical training purposes to optimize surgical techniques and learn how
surgical error leads to deviations from normal knee biomechanics.
The thesis will carry on with work which was already started before (Figure 1, [1]). A master thesis was also already devoted to this project last year
and concentrated on the design of a porcine knee joint mockup. This has shown that it is certainly feasible, though not easy, to create such a synthetic
surrogate of a porcine knee joint. The main objective of the thesis for next year is to finetune and further validate the result of this year's thesis and then
switch to a mockup of the human knee joint. Rather than exactly mimicking the complete structure of the knee, it will be sufficient to limit the device to
consist of the major bones, ligaments (collateral and cruciate ligaments), menisci and articular cartilage. Thus, it should be possible to arrive at a device
which displays similar kinematics and kinetics as the normal human knee during passive (unloaded) motion as well as weightbearing deep knee bends.
Work description:

As a first step, CT and MRI scans of five porcine knee joints will be made and their biomechanics (kinematics and 3D joint stability) will be fully
characterised and related to the mechanical properties of cartilage, ligaments and menisci (as measured using tensile and indentation tests). Based on
this information (as well as literature data), these five joints will be recreated using synthetic materials. As a validation, the native soft tissue will then
be replaced (one by one) by the artificial alternatives and the joints will be tested again and compared with their native behaviour.
Once this step has delivered acceptable results, the human knee joint will be the second target. A large amount of CT and MRI scans of normal human
knees is available. This information will provide the input to decide on the shape of the bones and the insertion sites of the main ligaments. The
materials which were identified in the first step, will then again be used to produce synthetic substitutes for the human ligaments and menisci. After
that, prototypes may be built and tested in terms of their kinematics (both passive and loaded) and 3D joint stability. The resulting data will then be
compared with similar data collected from the specimens that were also used for the CT and MRI scans.
References:
[1] Wong et al., “Validation of passive kinematics of a physical knee model with simulated soft tissues”, Transactions of the 56th Annual Meeting of the
Orthopaedic Research Society, March 6 – 9, 2010, New Orleans
Additional Info:
Materials testing of porcine and human tissues
An implantable system for bladder EMG measurements – #43
by Weydts Tristan – last modified Mar 29, 2017 15:31
Promotor: Bob Puers – Mentors: Tristan Weydts , Marko Bakula
Electromyography or EMG has a variety of clinical and biomedical applications. Measuring EMG signals originating from the bladder wall muscle is a
topic of ongoing research, as well as translating this task to a fully implantable system. There is a wide range of possible topic choices within this
thesis, such as: the choice of the electrode type, as well as their manufacturing, optimization of the measurement electronics, working on the lowpower
readout/communication system for the implanted EMG device and inductive powering of the system. The student can choose at the start which sub
problem they will focus on. The system should be biocompatible, so safety aspects must be considered and the device should be as small as possible
to ensure the comfort of the patient.
Additional Info:
Possible attendance of animal experiments (operating room).
Multimodal characterization of osteoarthritic cartilage – #44
by van Lenthe Harry – last modified Mar 29, 2017 17:06
Promotors: Davide Ruffoni (U. Liege) , Harry van Lenthe – Mentor: Harry van Lenthe
Osteoarthritis (OA) is a widespread degenerative joint disease that can affect any joint in the body. OA is characterized by a progressive degeneration of
the cartilage as well as of the underlying subchondral bone resulting in pain and limitation of joint movement. The main goals of OA treatment involve
reducing or eliminating pain and restoring joint mobility. Considering OA of the hands, there are several treatment options ranging from antiinflammatory
medications to (for the most severe scenarios) joint fusion, where the bones are fused together after the arthritic region is removed.
Despite numerous years of research, understanding the biomechanical properties of cartilage during OA is still a very active research topic. Cartilage is
traditionally divided into superficial, transitional, deep, and calcified zones. In contrast to the vast literature on the mechanical properties of superficial
and transitional cartilage, much less is known about the properties of deep and mineralized cartilage. The latter is, however, of high clinical relevance as
being responsible to connect cartilage to bone and being characterized by abnormal thickening during OA.
This Master Thesis, a collaboration between KU Leuven (Prof. van Lenthe) and ULg (Prof. Ruffoni), aims at characterizing the biomechanical properties
of OA cartilage with a focus on the region of calcified cartilage by combining image analysis based on microcomputed tomography and mechanical
testing based on nanoindentation. The specific workflow includes: i) sample preparation for microCT imaging ii) development of a protocol for microCT
imaging including phase contrast iii) sample preparation for nanoindentation and iv) nanoindentation testing. The main outcome of the Master Thesis will
be an optimized workflow for the biomechanical imaging and characterization of cartilage in OA.
MAIN TASKS
• Review the relevant literature on OA, focusing on OA of the hand
• Prepare OA sample for microCT and nanoindentation
• Perform microCT
• Perform imageguided nanoindentation testing
• Correlate microCT with nanoindentation measurement
• Write a detailed report and prepare a presentation of the work performed
PRACTICAL INFORMATION
• Project type: 80% experimental, 20% data analysis
• Project locations: Biomechanics Section, KU Leven and Mechanics of Biological and Bioinspired Materials, ULg.
• Project supervisor: Prof. Harry van Lenthe, Prof. Davide Ruffoni
• Required background: none
FE validation of head models through accident reconstruction – #45
by Woering Michel – last modified Mar 29, 2017 17:26
Promotor: Jos Vander Sloten – Mentors: Michel Woering , Markos Kapeliotis
Traumatic brain Injury (TBI) is one of the major health issues in this world and it affects people from all ages. Head injuries and TBI are the leading
cause of deaths amongst road users, with the highest numbers for pedestrians and cyclists. 70 to 90% of the deaths among cyclist is due to TBI. From
an ethical point of view it is hard to increase the understanding of TBI and other head injuries. Cadaver and animal experiments are done to increase the
knowledge, but less is known about the conversion of results to living humans. Therefore we use accident reconstructions from real life accidents.
Accident reconstruction uses rigid body models or FE models to reconstruct accidents. The process used in our group is based on fast rigid body
simulations with MADYMO (MAthematical DYnamic Models) followed by computational heavy FE simulations of the head in LS Dyna. Madymo is a
validated software package for accident reconstructions and will generate velocity and acceleration profiles as well as the impact place. A database
with cycling accident descriptions and medical information will be the base of this research. The accident description will be the input of the Madymo
simulations. Multiple crash scenarios will have to be explored to generate the best reconstruction results. Here the student will learn to work with
MADYMO and get experience with a fast mathematical simulation technique often used by car manufacturers.
Then, in order to get more insight into what happens inside the head a stateoftheart FE head model will be used. The acceleration profiles obtained
from the accident reconstructions will be used as an input. The outcome from these simulations could provide valuable insight in both further FE
validation of the numerical head model and more sophisticated helmet design. Here the student will have the opportunity to learn to work with FE
software the industry is using used for multiple purposes, including crash simulations.
In the end, the FE results can be compared with the CT/MRI scans of the accidents and the accuracy of the simulations can be calculated. This will
help to develop the FE head models and move accident reconstruction to the next level.
Objective:
The goal of the thesis is to investigate bicycle accidents, determine falling profiles and simulate the head impacts to determine their biomechanical
effects on the head.
Real life accident reconstruction with MADYMO
Work on realistic bicycle crash kinematics
Obtain velocity and acceleration profiles
FE Modelling
Simulate the head impacts
Analyze the outcome to validate the models results
This thesis gives the student the opportunity to work with rigid body modelling and finite element software in stateoftheart research on head impact
biomechanics. The results of this thesis are likely to lead to a publication.
More info...
Improving the preoperative treatment planning of distal radius fractures – #53
If you fracture your radius, the bone ends can heal in an abnormal position; clinically, this condition is called a malunion (Fig 1). The malunited radius
can cause complaints: pain, reduced mobility, and reduced strength. This can prevent patients to perform daily tasks and can cause early osteoarthritis
of the wrist. A corrective osteotomy (surgery) is often needed to correct the misalignment to normal anatomical position (11° 12° volar tilt).
Conventional treatment uses planning on tracing paper with standard X ray, during surgery K wire pins are used to transfer the preoperative plan into the
surgery. In general an opening wedge osteotomy is used. This type of osteotomy will create prominent free edges on the open side. These sharp, free
edges can cause problems with the extensor tendons.
More recently anatomical plates have been introduced to treat distal radius fractures. The design of these plates incorporates the creation of volar tilt.
Several authors have claimed that these anatomical plates allow the recreation of the appropriate anatomical relations, though in reality there will be a
variable amount of remodelling on the volar surface of the radius. The end result is often an undercorrection of the malunion when this technique is used.
This is the point where you, as engineer, can make a difference to help surgeons better treat his patients by optimizing the preoperative planning.
Aim 1: To develop and to optimize the preoperative planning based on 3D computed tomography images, using the contralateral (the other) intact arm
as the reconstructive goal.
Aim 2: Secondly, quantify the bone remodelling in the malunited radius as a function of time and fracture type.
Aim 3: Quantify the differences between the three reconstruction strategies and define the boundary conditions to evolve toward a (semi) automated
planning strategy.
This thesis gives the student the opportunity to work in close collaboration between engineers and an orthopaedic surgeon.
What is the effect of altering kinematics on the stresses and strains in the femoral and acetabular
cartilage? – #57
by Jonkers Ilse – last modified Mar 30, 2017 10:55
Promotors: Ilse Jonkers , Nele Famaey – Mentor: mariska wesseling (KU Leuven)
Musculoskeletal models are o擌�en used to analyse joint loading during human mo��on, such as gait. Using these models, the movement
pa��ern of a subject can be simulated based on mo��on analysis data. This way muscle forces that are required to perform the movement
can be calculated, as well as the contact forces in the joints of the lower limbs. Due to degenera��ve joint diseases (such as
osteoarthri��s), pa��ents adopt an altered gait pa��ern. Using a musculoskeletal modelling approach, the effect of this altered gait pa��ern
on joint loading has been studied and specific gait strategies favourable for reducing hip loading have been iden��fied.
Using finite element (FE) modelling, the effect of muscle and contact forces on stresses and strains in ��ssue, i.e. car��lage, can be
calculated. Typically, joint loading as measured in vivo in pa��ents with an instrumented hip prosthesis is used as boundary condi��on for
the FE analysis. However, this way subject‐specific joint loading is not included and therefore the effect of subject‐specific gait
characteris��cs on car��lage loading is omi��ed. This is highly important when assessing the effect of pathology, as previous research
already showed that subject‐specific muscle forces and joint loading affect the stresses and strains in bone, and should therefore be
taken into account when evalua��ng car��lage loading.
The goal of this porject is to create a FE model represen��ng the hip joint geometry, including the femoral and acetabular car��lage, and
to calculate stresses and strains in the car��lage for different gait pa��erns that were previously iden��fied to reduce joint loading. THis
study can then evaluate if these gait stratefies also op��mize the stress and strain distrubu��oin in the femoral and acetabular car��lage.
Additional Info:
This project will also expose you to the use of static and dynamic MRI
Can altered loading contribute to knee cartilage damage in OA patients? – #62
by Jonkers Ilse – last modified Mar 30, 2017 11:21
Promotors: Ilse Jonkers , renee van donkelaar (TU delft) , Nele Famaey – Mentor: maria pastrama (ku leuven)
We have previously used musculoskeletal models to quan��fy knee joint loading during various mo��ons, like gait and stair ascent and
descent in subjects with knee OA. Based on previous work, we found aberrant loading magnitude and loca��on, already from early
stages of the disease.
Within this thesis we want to take the next step and relate the aberrant loading to failure mechanisms of knee car��lage. This requires
first the calcula��on of the loading condi��ons in the ar��cular car��lage using an FE model (Openknee), with the loading condi��ons
obtained from the musculoskeletal model as boundary condi��ons. Therea擌�er, a workflow developed at the TU Del擌� will be customized
to relate collagen damage and proteoglycan deple��on to the mechanical loading profile in the knee ar��cular car��lage.
This way we will aim to es��mate the contribu��on of altered mechanical loading to car��lage damage in OA pa��ents.
What is expected of the student:
Use a detailed finite element model of the knee joint to calculate stresses and strains in cartilage for different motion patterns, using boundary
conditions calculated as with musculoskeletal modelling.
Customize and use an adaptive FE model to calculate the failure risk of collagen and proteoglycan depletion in the knee joint.
Additional Info:
collaborative research project with TU Delft
Development of a collagen network computational model for the modelling of cellextracellular
matrix mechanics – #50
by Heck Tommy – last modified Mar 30, 2017 11:34
Promotors: Hans Van Oosterwyck , bart Smeets (MeBios) – Mentors: Tommy Heck , Bart Smeets (MeBios)
Figure 1: A) Contractility in the actin cortex of the cell for an actively spreading cell on a rigid substrate. B) Computational model of a semiflexible
polymer network (red) enriched by molecular motors (green)
Project description
The interaction between cells and the extracellular matrix (ECM) plays an important role in many processes in the human body like tissue development,
tissue regeneration and angiogenesis. Cells embedded in the ECM receive mechanical and chemical stimuli from the ECM, degrade the ECM and apply
protrusive and contractile forces to it to migrate. In order to get a better understanding of mechanisms of cell migration, computational models are being
developed. Current models focus mainly on modeling of the cell, but often do not capture ECM mechanics and cellECM interaction. Therefore, the aim
of this thesis is to develop a computational model of mechanical cellECM interaction.
For the cell model, a model developed before for passive cell spreading will used (see Figure 1A) [1]. In this model, a triangulated mesh represents the
cell membrane and actin cortex. This model captures the physics of the actin cortex and allows for the localized calculation of contact and frictional
forces with the environment. So far, the model has been used for the modeling of cell migration on a flat substrate, but has not been embedded yet in a
3D model of the ECM.
For the ECM model, the student will develop a new model that captures the mechanics of discrete collagen fibers in a network. A computational model
has already been developed for the simulation of semiflexible polymers (capturing stiffness and bending) (see Figure 1B). By validating this model for
mechanical properties of collagen fibers and generating a crosslinked network of these fibers, a model representing a collagen network will be made.
Next, the deformable cell model will be embedded in order to investigate cellECM interaction. Results of this model will be compared with traction force
microscopy (TFM) data from experiments. This should lead to a better understanding of cellECM mechanical interaction for cells embedded in a
collagen network.
Both the cell model and the collagen model will be implemented in the C++/python based software platform called Mpacts (developed at division of
MeBioS at KU Leuven, see [3]). Having experience in programming is beneficial, but not required.

[1] Odenthal, T. et al., PLoS Computational Biology, 9(10):e1003267, 2013
[2] Pešek, J et al., Soft Matter 12:33603387, 2016
[3] http://demresearchgroup.com/
More info...
In vitro kinematics of medially arthritic knees: UKA and UKA combined with ACL reconstruction –
#67
Background:
Unicompartmental knee arthroplasty (UKA), in which only the arthritic parts of the cartilage are replaced, currently experiences increased popularity. It is
usually assumed that UKA leads to kinematics closer to the natural knee than total knee arthroplasty (in which the complete joint is replaced). Another
leading reason for the popularity of UKA is to preserve the majority of the soft tissues in the knee. Especially in younger patients, more natural knee
function and faster recovery have helped to increase the popularity of UKA. Many biomechanical questions remain, however, with respect to this type of
replacement. One of these is for instance whether it is also possible to perform a UKA in combination with a reconstruction of the anterior cruciate
ligament (ACL). Indeed, 25% of knees with medial compartment osteoarthritis also have a deficient ACL. Those patients would thus potentially benefit
from a combined UKA and ACL reconstruction.
To try to answer these questions, an in vitro cadaver knee experiment was performed in the past. Six cadaver specimens were prepared to be inserted
in a special knee kinematics simulator. The natural knees were then subjected to several loaded and unloaded motor tasks and their kinematics and
kinetics were measured. Then, the cartilage of the medial side of femur and tibia was removed to simulate osteoarthritis and the same tests and
measurements were repeated. In a third step, a UKA was performed, after which the same tests and measurements were repeated again. Finally, the
ACL was cut and reconstructed and the tests and measurements were redone once more.
The first goal of the master thesis is to analyse the collected data and compare (also statistically) the behaviour of the knee joint after each step in the
experimental procedure. Software for this purpose is largely available, although a limited amount of new software will also need to be written in Matlab.
After the analysis of the experimental data is done, a musculoskeletal model of the experiment will be developed (in OpenSim). The goal is to try to
recreate the experimental results and eventually perform a more in depth analysis of the altered knee behaviour.
Design of a muscle module to augment an existing robotic gait simulator – #68
Background
Mobilab disposes of an industrial robot which is used to simulate gait with artificial lower legs. The robot end effector follows the trajectory of the knee
joint (as measured in vivo in the gait lab) and holds the lower leg in the right orientation. At the same time, the correct ground reaction force as a
function of time is also obtained. Thus, the robotic gait simulator can be used to investigate, in vitro, the performance of different alignments of upper
and lowerleg prostheses, the effect of different shoe types, … . It can also be used to perform tests which are not always feasible in vivo (due to the
risk of falling of a patient, for instance). However, the robot cannot be used to simulate situations where a real, actuated, knee or ankle joint is still
present.
Therefore, we want to add an actuated ankle and/or knee joint to the robot to enhance its applicability and make it also useful for the field of
orthopaedics.
The goal of this design project is to design a module which can be added to the robotic gait simulator and which consists of artificial muscles
(pneumatic, hydraulic, electromagnetic,…actuators) that can be controlled based on position of the robot endeffector and/or ground reaction force
values. The correct muscle force to apply at any instant of the gait cycle (and other motor tasks) will be derived from musculoskeletal models,
experiments or literature.
Detailed description
Two separate modules will be designed, one for the knee and one for the ankle (but keeping in mind that it should also be possible to combine them).
Based on muscle forces in the knee and ankle for different motor tasks (gait, chair rise, stairs, squat), on control and precision and on weight limits for
the module, a selection of the right actuators will be made. Then, a mechanical design to couple the actuators and the robot will be developed as well as
a system to communicate with the robot and control the actuators based on the input signals from the robot and the force plate.
Investigating the potential of µCT imaging for cardiovascular tissue – #69
by Famaey Nele – last modified Mar 30, 2017 14:07
Promotors: Nele Famaey , Harry van Lenthe – Mentors: Markos Kapeliotis , Emma Vanderveken (KU Leuven)
Background
Recently, a stateoftheart µCT scanner was acquired and installed at BMe, as part of the new Leuven Biomechanics Testing Facility. This device is
currently already used for the full 3D microscale characterization of bone. Reconstruction of the bony tissue is possible due to the high contrast
between bone and environment. For soft tissues, this contrast is far less pronounced, which is why soft tissues have been very difficult to characterize
using µCT. However, the new device also has an innovative feature called ‘phasecontrastimaging’ (see http://bruker
microct.com/next/SkyScan1272brochure.pdf for more info). This technique should allow a more detailed differentiation, making it possible to also
identify soft tissue. Moreover, the device is equipped with a ‘materials testing stage’, which will allow for the testing of samples under load (tension or
compression).
Objectives
The aim of this thesis is to explore the functionalities of the new µCT scanner for soft tissue characterization, and more specifically to find the optimal
conditions and settings for cardiovascular tissue imaging. Different types of cardiovascular tissue samples (human aorta, bridging vein, mitral valve) will
be available to this end. The aim is to obtain a full 3D characterization to the highest possible degree of accuracy, and at different levels of tension. If
time allows, the obtained 3D geometries can be converted into finite element simulations for inverse material parameter optimization.
Main tasks:
Finding the optimal settings of the scanner for various cardiovascular tissue samples.
Investigating different contrast agents for enhanced contrast (e.g. exitron, sodium polytungstate).
Experimenting with the materials testing stage for imaging sample under tensile load.
Converting the obtained 3D volumes into finite element meshes.
Note that, although strong support and guidance will be provided on a regular basis, a high level of inquisitiveness and autonomy will be required in this
thesis, as many of the techniques are still unknown even to your mentors and supervisors.
Additional Info:
Automatic segmentation of 4D CT images – #72
by Boey Hannelore – last modified Mar 30, 2017 14:43
Promotors: Jos Vander Sloten , Ilse Jonkers – Mentors: Hannelore Boey , Christophe Van Dijck
The understanding of anklefoot biomechanics is lagging behind, compared to the other lower limb joints. Given it’s complex structure and the need to
study the biomechanics during weight bearing, the use of medical imaging based approaches have not been fully exploited. For one of our ongoing
projects, we have developed a unique foot manipulator to be used in a 4D CT scanner, to measure individual foot bone kinematics during imposed gait
movements. The ability to measure the individual foot bone kinematics is important as such information can be used in a wide range of applications:
detecting joint instabilities, helping in preoperative planning and postoperative evaluation, studying the effect of joint degeneration and comparing
different treatment methods.
During the data acquisition, the subjects are laying supine in the 4D CT scanner with the foot attached to the foot plate of the manipulator that imposes
a walking movement to the foot while loading the foot. During a movement of 10s, 37 CT scans are acquired using a specific dynamic CT sequence.
This results in large dataset of CT images that need to be segmented to retrieve the position of the individual foot bones. At the moment, we are using
an automatic segmentation code that is developed by a PhD student at the Biomechanics Division (Christophe Van Dijck). Within the current thesis, we
aim to increase the accuracy of the segmentation process and to incorporate time dependent features in the imaging sequence to allow a more accurate
and faster determination of the individual foot bone position during the imposed motion.
This thesis is integrated in the footankle group which is a collaboration between the division ‘Biomechanics’ (Faculty of Engineering) and the ‘Human
Movement Biomechanics Group’ (Faculty of Kinesiology and Physiotherapy).
2. Biomedical Data Processing
Mining public health data records using various machine learning methods – #4
by Tassens Ida – last modified Mar 22, 2017 12:12
Discipline: Biomedical Data Processing
Promotor: Bart De Moor – Mentor: Gorana Nikolic
INTEGO database contains over 3.5 million diagnoses, 36 million lab results, 15 million medication prescriptions and that makes it perfect for data
exploration. The data is used for various studies from incidence and prevalence rates, epidemiological studies, finding comorbidities and etc. The data
is mostly observational data gathered from the GPs clinical practitioners. This makes it interesting for a number of statistical methods – from basic to
advanced ones
The aim of this proposal is to explore the use of machine learning methods on the INTEGO data. We have various applications, from generating novel
hypothesis on comorbidities to, using hidden markov models to describe unobserved processes in clinical data, that will be formulated in close
collaboration with GPs from the ACGP.

Additional Info:
Kind of work : Literature 35 % Programming 65 %
Profile : good knowledge on Python/R programming and basic knowledge of the sckitlearn python library

Mining Health Care Records – #3
Promotor: Bart De Moor – Mentors: Gorana Nikolic , Marc Claesen
The INTEGO database of the Academic Centre for General Practice (ACGP) at KU Leuven contains EHR (Electronic Health Record) data registered by
ca. 100 Flemish GPs since 1994. At the moment, the database contains information on ca. 380.000 patients. The INTEGO patient population is
representative for the Flemish population with regards to age, gender and socialeconomic status at municipality level. The database contains over 3.5
mil diagnoses, 36 mil lab results and 15 mil medication prescriptions and.

The database is used for epidemiological studies on incidence and prevalence rates, distribution factors etc. To do that, the data are considered as
coming from cohort studies, but these are purely observational data that were gathered as part of a GPs clinical practice, not as part of a research
study. As such, the data pose a number of challenges to the use of classical statistical methods.

In this thesis, we aim to explore the use of machine learning methods on the INTEGO data, and visualize them using different tools (i.e. d3.js).
Applications can range from generating novel hypotheses on occurrence of comorbidities to studying the effect of comedications, and will be
formulated in close collaboration with GPs from the ACGP.

Additional Info:
This project assumes a background in computer science or mathematical engineering and a special interest in health care applications

The brain as a complex system : Agentbased modeling of neurons – #5
Promotor: Bart De Moor – Mentor: Supinya Piampongsant
Context
The study of how the human brain works has long fascinated scientists. Within the last decades, dramatic improvements in data acquisition techniques
and computational power have led to the development and application of novel mathematical tools for the study of brain complexity. One of these
includes the framing of the brain as a complex system. Methods in network science and systems theory are being combined to form a new
interdisciplinary discipline for the study of systems with large numbers of components and complex interaction rules.
Goal
The goal of this thesis is to model the brain as a complex system. The student will work with mathematical models at multiple spatial and temporal
scales, starting from the neuronal level, building up to parts of the brain, and to the whole brain itself, time permitting. Multiple combinations of neuronal
models and their interaction rules will be tried. The resulting system stability, complexity, efficiency, costs and emergent behavior will be examined and
compared.

Additional Info:
Kind of Work
Literature and analysis (40%) / programing and simulation (60%)
Profile (e.g. rather theoretical / rather practical implementation, courses, methods, computer language(s) etc.)
Good knowledge of biological systems and understanding of (at least to the level outlined in Steven H. Strogatz’s “Nonlinear Dynamics and
Chaos”)
Proficiency with MATLAB and Simulink
Basic knowledge of optimization, parameter estimation
Theoretical knowledge of network topography
Strong interest and willingness to take initiative. The student can be given some freedom in shaping the direction of the project.

Constraintbased reconstruction and analysis of gene regulatory network – #6
Promotor: Bart De Moor – Mentor: Supinya Piampongsant
Context
A genetic regulatory network functions as an interplay of numerous interacting biochemical components. Understanding how a genetic network works
can be crucial for understanding diseases and discovering cures. However, it is not always possible for a biologist to take measurements of every
component in a genetic network, thus rendering many parameters in the network unknowable. To solve this problem, and to help elucidate the underlying
structure of a genetic network, various methods have been attempted.
Goal
In this thesis, we explore constraintbased network reconstruction and analysis. The method, already popular in the field of genetic engineering, seeks
to limit a network’s possible topologies according to its physiochemical, environmental, spatial, and selfregulatory constraints. The student will be
applying flux balance analysis to public data in an attempt to reconstruct a metabolomics network. S/he will compare the results to that in existing
literature and draw conclusions regarding the effectiveness of this method compared to others.
KIND
Additional Info:
Kind of work
Literature and analysis (20%) / Data collection and preprocessing (20%)/Learning and training technical skills (30%) programing and simulation (30%)
Profile (e.g. rather theoretical / rather practical implementation, courses, methods, computer language(s) etc.)
Be able to work and gain new skills independently
Basic working knowledge of optimization, parameter estimation, and MATLAB/Simulink programing
Familiarity with COBRA toolbox a plus
Strong interest and willingness to take initiative. The student can be given some freedom in shaping the direction of the project.
How to deal with missing values in urinary metabolomics from NMR spectroscopy ? Impact of
imputation schemes – #7
Promotor: Bart De Moor – Mentor: Thibaut Vaulet
Missing values is a common problem for metabolomics data. Around 1040 % of a metabolomics dataset usually contain missing values. However,
some machine learning algorithms used to subsequently analyzed metabolomics data cannot handle missing values. Therefore, it is a common practice
to impute the missing values.
The origin of those missing values can be caused by three principal reasons: 1. the metabolite is not present in the sample; 2. the metabolite is present
but is missed during preprocessing; 3. the metabolite is present in the sample but is below the level of detection.
Common strategies to deal with missing values used in metabolomics data are marginalization, mean or zero imputation or setting all missing values to
half of the minimal value present (i.e. assuming the missing value are missing because the concentration is below the detection threshold). However,
those approaches are rather naïve and other strategies using as kNN or RF imputation have shown excellent results.
It has been demonstrated that depending on the strategy chosen to handle the missing values, outputs of subsequent data analysis can be drastically
different. Therefore, it is of crucial importance to carefully select the right approach beforehand.
In this project, the student will implement different strategies to deal with missing values in metabolomics data (artificial and real data) and subsequently
benchmark the performance of different classifiers (PLSDA, PCDFA, SVM, RF, ...)

Additional Info:
Required skills : Knowledge in missing data analysis and machine learning, good programming skills (python ,R)

Automated neonatal seizure detection using Deep Neural Networks – #23
by Ansari Amir Hossein – last modified Mar 28, 2017 13:52
Promotors: Sabine Van Huffel , Gunnar Naulaers – Mentor: Amir Hossein Ansari
Background:
Automated neonatal seizure detector (ANSD) is a fundamental processing unit of an automated system in neonatal intensive care units (NICUs). An
interdisciplinary team at KU Leuven has developed a sophisticated system, NeoGuard, to monitor the newborn babies in the NICUs. The current
seizure detection method was first time published in 2008, and has been improved in 2010, 2011, and 2016. This method is one of the most accurate
ones in the literature. This heuristic method has some advantages like high accuracy and transparency. However, the main disadvantage is that it is a
fixed detector and is not retrainable. Therefore, we are looking for a new datadriven method which is high accurate, retrainable, and has preferably
interpretable structure.
Deep neural network (DNN) is one of the wellknown techniques in image processing and speech recognition problems. There are dozens of different
types of deep neural networks like convolutional neural networks (CNN), deep belief networks (DBN), stacked auto encoders (SAE), long short term
memories (LSTM), and etc. It seems that the unique characteristics of CNN, which is currently the stateoftheart and the most common technique in
imageprocessing, can be wellfitted in seizure detection problem.
Goal and Objectives:
Recently a deep CNN was designed in our group and tested on some neonates. Although the results are promising and better than other implemented
datadriven methods, it is still weaker than our heuristic algorithm. The main goal of this master project is redesigning the current network or designing
other deep networks, like DBN or SAE, in order to improve the performance of neonatal seizure detection problem. It is expected that other simple
featurebased classifiers are also made by the student and their outputs are compared with the outputs of the DNN to show its added value.
The prospective student will be guided by the biomedical signal processing researchers in a close collaboration with the NICU of UZ Leuven,
Gasthuisberg. Furthermore, a regular meeting about neonatal monitoring in the hospital provides a unique opportunity for the student to hear about in
progress studies in this field, present his/her ideas or results, ask his/her questions, and use the medical doctors’ feedback.
Additional Info:
Required Skills:
Matlab experience (required),
Biomedical signal processing (required),
Patternrecognition (required),
Recommendations:
Since the DNNs are working much faster on GPU than on CPU, it is seriously recommended that the student has a laptop or pc with nVidia
GPU.
We have already a few Matlab toolboxes for deep learning. But there are many powerful python toolboxes implemented by Google, Microsoft and
other famous research teams in this field. If the student knows python programming language, it can be very useful.
Take up the challenge of personalized Medicine of Diabetes by Machine Learning and Big Data
analytics with Imaging Mass Spectrometry – #8
Promotor: Bart De Moor – Mentor: Xian Mao
Diabetes and pancreas
Diabetes directly causes about 1.5 million deaths each year worldwide and is diagnosed in 8~9% of adults in all income levels. The global agreement of
diabetes is to effectively control its rise by 2025. And pancreas is fundamental to this goal since it's the workstation of insulin.
Many biomedical studies are under an assumption that their samples are of homogeneity. But this assumption is flawed and subsequently it can
produce unexpected outcomes that are harmful to patients. We now with the help of IMS can zoom in to investigate the heterogeneity
in details.

Imaging Mass Spectrometry (IMS) and Big Data
As the The Nobel Prize in Chemistry 2014 also highlights the contribution of molecular imaging in life science, IMS is a rising technology that rapidly
produces images of each of the tremendous amount of molecular species in the sample tissue[ Figure 1]. By the virtue of these
labelfree and informative characteristics, IMS attracts growing interests from various fields such as biomedicine, pharmaceutics, fundamental biology,
forensics and etc.. And the size of IMS datasets ranges from Gbs to TBs. To interpret and exploit these big data, modern machine
learning algorithms plays a crucial role as well as computing platforms dedicated to big data analytics including clusters, GPUs, clouds and
neuromorphic.

Drug discovery
Discovery of satisfactory drugs has been always an appeal from patients, health organizations and companies since this is a very tedious process.
Particularly, the effects of drug candidates are not always all predictable and the targets often hard to find and hard to validate. On the other hand, IMS
can amend this situation because its high resolutions in space and m/z provide holistic and actionable information of the metabolism of the tissue.

Approaches*
To study the effect of the drug candidate via big data analytics and machine learning such as Genetic algorithms and Deep learning, we'll model the
changes of drug candidates, insulin and glucagon in terms of abundance and distribution pattern. We can then evaluate its therapeutic
and potential side effects by computing the aberration of metabolic profiles. Meanwhile, by analyzing the metabolic IMS profiles of our control
experiment, we will look for other molecules of interest for example ubiquitin, lipid regulators, splice variants and molecules having
significantly increased or decreased abundances with the help of databases and networks.
Next we will use these molecules, served as the source of reverse translation, to generate new hypothesis for target and biomarker discovery.
Additionally, we will employ recent advancement in Visual analytics for conducting analytical reasoning in an interactive way.
Additional Info:
*The suggested approaches can be carried on independently in parallel by multiple students. Skills and experience learned from this work are flexible
and transferable to other tasks in big data analytics. PhD positions are possible after this work.

Literature (34 %), Programming (33%), Writing (33%).

Academic activities
This work is in close collaboration with the Faculty of Medicine plus the external academic and industrial parties. And we aim to publish relevant
outcomes in the peerreview journal.

Requirements
Basics of technical computing (e.g. MATLAB, Julia, R, python, etc.) ; Courses in Machine Learning
Evaluation of statistic characters for imaging mass spectrometry data – #9
Promotor: Bart De Moor – Mentor: Xian Mao
Imaging Mass Spectrometry (IMS) and Big Data
As the The Nobel Prize in Chemistry 2014 also highlights the contribution of molecular imaging in life science, IMS is a rising technology that rapidly
produces images of each of the tremendous amount of molecular species in the sample tissue[ Figure 1]. By the virtue of these
labelfree and informative characteristics, IMS attracts growing interests from various fields such as biomedicine, pharmaceutics, fundamental biology,
forensics and etc.. Additionally, the size of IMS datasets ranges from Gbs to TBs. To interpret and exploit these big data, modern
machine learning algorithms plays a crucial role as well as computing platforms dedicated to big data analytics including clusters, GPUs, clouds and
neuromorphic.

Drug discovery
Discovery of satisfactory drugs has been always an appeal from patients, health organizations and companies since this is a very tedious process.
Particularly, the effects of drug candidates are not always all predictable and the targets often hard to find and hard to validate. On the other hand, IMS
can amend this situation because its high resolutions in space and m/z provide holistic and actionable information of the metabolism of the tissue.

Approaches*
Due to this nature of high dimensionality and complexity of IMS datasets, it is crucial to ensure that we extract knowledge of genuine biomedical
relevance rather than artifacts fabricated by diverse groups of noise and debatable analysis. So proper preprocessing is apparently a
determinant for any following analysis of the valuable and informative dataset. However currently the need for the proper preprocessing hasn't been met.
Thus it is important to investigate the intrinsic characters of these data in terms of Statistics as the primary step. To achieve this, we will use classical
and novel methods of descriptive Statistics, probability distribution and hypothesis tests. Then based on these analysis, we will determine which steps
in preprocessing are needed and what are the specifications. For instances, one of the steps is the normalization of spectra. Akin to normalizations in
many other scientific and engineering fields, this normalization step is aiming to make all the spectra into one scale in order to allow justifiable
comparison among spectra within one dataset and between multiple datasets. It is the justifiable comparison that makes the subsequent analytics such
as clustering and classification meaningful and useful for translational medicine. Nowadays, because of improper normalization, the clustering and
classification techniques implicitly are based on a false assumption that everywhere in the tissue shares very similar biochemical composition. So the
resultant discoveries are with questionable clinical value. For this task a dedicated artificial
tissue will be synthesized and its imaging mass spectrometry (IMS) data is going to be used to benchmark the performance of the statistic analysis and
algorithms of preprocessing. A real IMS dataset of cancer study will be tested using the best approach based on the benchmarking step. Moreover, we
will try to incorporate additional information to improve the results. In brief to foster the application of IMS in clinics, we will evaluate various
preprocessing approaches and improve this step for beneficial insights in biomedicine.
Additional Info:
*The suggested approaches can be carried on independently in parallel by multiple students. Skills and experience learned from this work are flexible
and transferable to other tasks in big data analytics. PhD positions are possible after this work.

Literature (34 %), Programming (33%), Writing (33%).
Spike sorting algorithms for separation of motorunit action potentials in highdensity surfaceEMG
– #15
by Wouters Jasper – last modified Mar 27, 2017 16:58
Promotor: Alexander Bertrand – Mentors: Jasper Wouters , Boudewijn Sleutjes (UMC Utrecht/Neurology division)
Neuromuscular disorders (e.g. ALS, mul��focal motor neuropathy, etc.) directly or indirectly affect so‐called motor units (MUs), which are the smallest func��onal
element of the peripheral nervous system that controls skeletal muscles. A MU consists of three basic components: an alpha‐motor neuron (lower motor neuron) in the
brain stem or spinal cord, an axon, and all the muscle fibers it innervates. As MU size, MU shape and also MU ac��vity may be affected due to pathology, their analysis
can be of great assistance in the clinical evalua��on of neuromuscular disorders. Besides clinical relevance, the analysis of MU ac��vity is a basis for human‐machine
interfacing, especially for the control of advanced ac��ve prosthe��cs and orthoses (eg. exoskeletons).
A rou��nely applied electrophysiological technique in the diagnosis and follow‐up of neuromuscular disorders is the use of surface electromyography (sEMG), which
allows to measure electrical ac��vity from the muscle fibers in a non‐invasive fashion by a��aching a few electrodes on the skin above the muscle. The sEMG signal of an
MU is the summa��on of all muscle fibers that make up the MU, which results in a so‐called MU ac��on poten��al (MUAP), which is a signal consis��ng of a train of spiky
waveforms. However, o擌�en mul��ple MUs are simultaneously ac��ve, such that the sEMG signal will consist of a mixture of several MUAPs (+noise).To extract
individual MUAPs or spikes from this data, so‐called ‘spike sor��ng’ algorithms have been developed, which demix the MUAPs and clusters them according to their
underlying MU.


Recently, the use of high‐density sEMG HD‐sEMG) with a large densely spaced array of more than 100 electrodes has become a popular tool for MU analysis, as it
allows recording of not only temporal informa��on, but also spa��al informa��on of muscle ac��vity. As a result, it allows collec��on of a larger sample of different MUAPs
than conven��onal sEMG , and aids the recogni��on of single MUAPs. In this thesis, we aim to exploit the high spa��o‐temporal structure of HD‐sEMG signals for spike
sor��ng, i.e., we aim to combine the informa��on in all channels to extract MUAPs.
The goal of this thesis is toassess the limits of exis��ng spike sor��ng algorithms on HD‐sEMG data, and to come up with new/modified approaches to improve the
‘spike yield’, i.e., to extract more MUAPs. This would allow researchers and medical doctors to more reliably evaluate their abundance and firing pa��ern, which are
expected to be affected due to underlying pathophysiological changes. Hence, a detailed spike train analysis of MUAPs may add new and relevant clinical informa��on.
Also, more precise spike train analysis of mul��ple MUAPS enables be��er control of ac��ve prosthe��cs/orthoses.
Project
A擌�er a brief literature study, we will first assess exis��ng spike sor��ng algorithms on HD‐sEMG recordings from the University Medical Centre Utrecht and Erasmus
University Medical Centre Ro��erdam.These spike sor��ng algorithms will typically first detect spikes in the signals, extract features from them, and then use clustering
tools to find clusters of spikes corresponding to the same MU, finally a template matching algorithm resolves overlapping spikes.
In the second stage, our goal is to improve the spike yield by combining the strengths of theseexis��ng algorithms, and extend them with other mul��‐channel signal
processing techniques (eg. based on mul��‐channel Wiener filtering, convolu��on kernel compensa��on, independent component analysis, etc.).
The programming is done in Python or MATLAB. For more informa��on, don't hesitate to contact us!
Promotor:
Prof. Alexander Bertrand (KU Leuven, ESAT‐STADIUS) ([email protected])
Supervision:
Jasper Wouters (KU Leuven, ESAT‐STADIUS) ([email protected])
Dr. Boudewijn Sleutjes (UMC Utrecht/Neurology division) ([email protected])
Work load: literature 20% ‐ implementa��on 45%‐ valida��on 35%
Development of a Brain Computer Interface based on the assessment of the common dynamics
between EEG and fNIRS recordings – #16
by Van Eyndhoven Simon – last modified Mar 29, 2017 10:07
Promotors: Sabine Van Huffel , Alexander Caicedo Dorado , Borbála Hunyadi – Mentor: Simon Van Eyndhoven
Functional nearinfrared spectroscopy (fNIRS) is a noninvasive technique that allows a noninvasive assessment of the changes in oxygenation of the
brain, which are the result of metabolic processes. As such, it provides information that is similar to the one obtained by functional MRI, but it offers a
higher sampling rate at the cost of a lower spatial resolution, and more importantly, can be integrated in a wearable device instead of a bulky scanner.
By merging information from the electroencephalogram (EEG) and fNIRS it is then possible to identify regions of high neural activity. This allows the
development of brain computer interfaces (BCI) which interpret and exploit these patterns of activity, which is of considerable interest in medical and
commercial applications.
In this thesis, we will explore the use of data fusion techniques in order to characterize the common dynamics between brain function, assessed by
means of EEG measurements, and changes in brain hemodynamics, assessed by means of fNIRS. We investigate if this information is useful in the
localization of brain regions that are active during a certain task, as well as to predict the type of task a person is performing. Datadriven methods
based on blind source separation, information dynamics, dynamic time warping, and kernel methods, among others will be explored. The data used for
this project has been used for the training of simple BCI classifier and is available from an open access data repository. Hence, the aim in this thesis is
twofold: extracting meaningful information from the complementary, multimodal data, and using these new features to improve the performance of
newly developed machine learning algorithms (with respect to the reference classifier).
Additional Info:
Prerequisites:
The student should be familiar with (biomedical) data processing and should have good Matlab skills (in other words, successfully having
completed course H03I2A is required).
Previous experience with fNIRS and/or EEG data is not required, but can be an asset.
More info...
Sleep stage classification in preterm infants using deep neural networks – #21
by De Wel Ofelie – last modified Mar 28, 2017 11:06
Promotors: Sabine Van Huffel , Gunnar Naulaers (UZ Leuven) , Alexander Caicedo Dorado – Mentors: Amir Hossein Ansari , Ofelie De Wel
Background
Babies born before 37 weeks of gestation are preterm and have an increased risk of neurological abnormalities.In the neonatal intensive care unit(NICU)
the brain maturation of these vulnerable babies are closely monitored by analyzing their EEG.
Neonatal sleep can be divided in 3 stages: awake, active sleep and quiet sleep. Monitoring the sleepwake cycling of preterm infants is important, since
it provides an indication of their brain development and maturation. Common practice to identify the different sleep stages is visual analysis of the
neonatal video EEG, but this is a time consuming and tedious task for the neurophysiologist. Therefore, an objective and automatic classification of the
different sleep stages can be of great help for the clinicians.
Several algorithms have been developed for neonatal sleep stage classification and they typically consist of 3 steps: preprocessing, feature extraction
and classification. Deep neural networks (DNN) are commonly used in image processing and speech recognition problems. In this project, we want to
apply DNNs in order to overcome the feature extraction step.
Goal
The aim of this thesis is to develop an algorithm for automatic detection of quiet sleep in preterm infants. For this purpose, the student will design
DNNs, such as convolutional neural networks, stacked auto encoders or deep belief networks. Afterwards, the classification performance will be
compared to other existing algorithms.
The research will be performed in close collaboration with professor Gunnar Naulaers, head of the Neonatal Intensive Care Unit, and his team of UZ
Leuven.
Result
The result of this master thesis is a rigorous algorithm that automatically detects the different neonatal sleep stages.
Additional Info:
The student should have some experience in writing Matlab code and should have followed the course “Biomedical Data Processing (BKULH03I2A)”.
Moreover, it is advised to have a pc with nVidia GPU since DNNs are working much faster on GPU. At last, knowledge of python programming
language can be useful.
Epileptic seizure detection at home using accelerometers: Taking it one step further – #24
by De Cooman Thomas – last modified Mar 29, 2017 16:58
Promotors: Sabine Van Huffel , Lieven Lagae – Mentor: Thomas De Cooman
Epilepsy is a neurological disorder that affects around 1% of the people worldwide. Around 35% of them cannot be cured by antiepileptic drugs or
surgery and have to live with their disease for the rest of their lives. The most vulnerable group of them are children with epilepsy. Not only the patients,
but also their parents are facing hard times. They typically want to be able to help their child whenever he or she has a seizure, but they cannot always
be near them, for instance during the night. Monitoring of these patients in the home environment can help to improve their quality of life. A warning
system for example could try to detect seizures and warn the parents in case of a seizure. Because EEG cannot be obtained easily outside the
hospital, other biomedical signals need to be used for this seizure detection system. Accelerometers for example can be used to detect strong
convulsive seizures, and are easily acquired in normal life situations (available in most smart watches).
Although already some research has been performed on this topic, a lot of questions remain about the automated seizure detection using
accelerometers. The goal of this master thesis is to give answers to some of these remaining questions through testing several possibilities for seizure
detection algorithms. Some potential questions that can be answered:
Extracting features from accelerometers can be done in two ways: feature extraction from the different channels inside a sensor (x, y, z) separately,
or feature extraction per sensor. Which option gives the best results?
How many sensors are sufficient for accurate seizure detection?
Can a oneclass classifier be sufficiently accurate for seizure detection?
Does including patientspecific data points result in a significant performance improvement?
A dataset coming from the Pulderbos revalidation center, a recognized center in Belgium specialized in the treatment of children with convulsions and
epileptic seizures, will be made available for this master thesis. In this dataset, the patients are wearing 4 accelerometers, one on each ankle/wrist. The
thesis will be in strong collaboration with child neurologists from both Pulderbos revalidation center and the child neurology department from UZ Leuven.
Additional Info:
Candidates should have a basic knowledge of machine learning concepts (e.g. crossvalidation, classifiers such as (ls)svm,...) or at least show interest
in them to learn about it. Basic programming skills in Matlab are required.
The thesis will be in strong collaboration with the UZ Leuven hospital under supervision of prof. dr. Lieven Lagae (child neurologist at the UZ Leuven).
Opening the black box in machine learning: the potential role of Interval Coded Scoring for
interpretable classification of clinical states in rheumatic diseases – #20
by Billiet Lieven – last modified Mar 30, 2017 14:31
Promotors: Sabine Van Huffel , Kurt de Vlam (UZ Leuven) – Mentors: Thijs Swinnen (UZ Leuven) , Lieven Billiet
Axial spondyloarthritis (axSpA) is a widespread chronic musculoskeletal disease. Its symptoms include low back pain, inflammation of the spinal
region, ossification of the spine (fused vertebrae) and sometimes also pain in the peripheral limbs. In current clinical practice measurements such as
MRI and blood samples are combined to assess the disease severity and guide the treatment process.
With the evolution of data science techniques, new information can be extracted from complex data. This has among other applications potential
benefits for medical practice. However, standard data science and machine learning techniques often focus on mere correct recognition or diagnosis
without offering a thorough explanation as to how this decision was obtained. Also, in the field of rheumatology, scoring systems such as ASDAS
have been implemented for patient followup. Typically, health care practioners combine patientreported questionnaires, clinical measures (e.g.
spinal mobility), radiology (MRI) and serology to estimate disease activity and other clinical outcomes. Unfortunately, such scoring systems may
contain bias, since they were mostly developed using expert opinion and provide only posthoc validation from measured data.
The Interval Coded Scoring (ICS) starts in a machine learning context, but aims to bridge the gap to the medical practice by providing interpretable
systems rather than black box solutions. More particularly, we developed a method based on sparse optimization to extract scoring systems from
data in a semiautomatic fashion. The ICS outcome differs from known scoring systems in several ways. For many scoring systems currently in use,
building the model is based on the variance of the data to indicate importance of variables. Secondly, current scoring systems are often linear
combinations of variables, disregarding the fact that the importance of a variable might change nonlinearly with their value. Furthermore, variable
interactions are rarely considered. All these limitations can be addressed with standard machine learning techniques such as Support Vector Machines,
but at the cost of interpretability. In contrast, ICS selects variables based on optimization, that is, directly based on their relevance to the task at hand;
it assigns weights reflecting the importance of relevant variable intervals, hence allowing nonlinearity; it considers variable interactions. Yet, the
outcome still remains an interpretable scoring system. An example of it can be seen in the figure on the right. It shows a model to diagnose vertebral
afflictions.
For this master thesis, several large and highdimensional cohort datasets have been acquired. These clinical datasets recorded groups of patients
suffering from axial spondylitis with various degrees of severity. In the treatment, the use of antiinflammatory TNF inhibitors is considered an indication
of the severity of the disease. Hence, it will be used as classification outcome for the datasets.
The thesis aims to further process the data, design a rigorous feature selection protocol and combine it with an interpretable classifier such as ICS to
obtain a clinically useful scoring system. All of this will be performed while remaining in close contact and discussion with clinical experts.
Next, the resulting system will have to be compared both to the outcome of current blackbox techniques and, more importantly, to the approach and
outcome of the current clinical practice.
Additional Info:
Since this is a real research case study, a strong researchminded disposition and some knowledge of basic machine learning is required, as well as
programming proficiency in Matlab.
EEGbased sound quality assessment: sound perception and its effect on human emotion – #35
by Das Neetha – last modified Mar 28, 2017 16:36
Promotor: Alexander Bertrand – Mentor: Abhijith Mundanad (ESAT)
Exposing a person to a specific sound for a long ��me can cause annoyance, fa��gue, or even stress, e.g., during a plane flight or when working long hours in a noisy
environment such as a machine room. Many companies, par��cularly in the automo��ve or machine industry, rely on sound engineers who have the difficult task to make
the sound of machines, engines or consumer products as pleasant as possible. In many cases it is not as straigh򦩔orward as simply reducing decibels, but also to make
the sound more pleasant to listen to (think about a Ferrari engine vs. a lawn mower).
However, sound percep��on is inherently subjec��ve. It depends on the volume, the strength of different frequencies, their modula��ons or varia��ons, and several other
characteris��cs, which are perceived differently by different subjects, and invoke different emo��ons. Therefore it is important to understand which features of sound
impact its nega��ve or posi��ve percep��on. Sound quality assessment is typically carried out by asking the opinion of mul��ple test subjects during what is called a ‘Jury
Test’, which is highly affected by several confounds, such as honesty, par��cipa��on effort, intelligence, preoccupa��on, etc. of the test subjects.

But, what if we can supplement the subjects ra��ng (opinion) of sounds by objec��ve measures of percep��on derived from neural responses in the brain? It has been
found that emo��onal states of pleasantness and sa��sfac��on can be gleaned from EEG signals of humans. These studies have even evolved to improving the
emo��onal state through specific sound s��muli or music. What if we can learn from these approaches and find a way to objec��vely assess sound quality/percep��on
during jury tests? It would lead to quicker and more reliable tests, be��er products, and a be��er understanding of how human emo��onal states are influenced by every
day sounds. In the long run, this could even help in assessing how and which sounds cause stress or are a health risk.

This thesis will be in close collabora��on with Siemens, one of the biggest electrical engineering companies in the world. The local Leuven branch cons��tutes the
headquarters of the Siemens “Simula��on and Tes��ng Solu��ons” (STS)1 business segment. Siemens STS is one of the world players on sound quality assessment2, in
par��cular for the automo��ve industry, where they collaborate with design engineers of Ferrari, Mercedes, etc.

Your work: In the area of sound quality assessment, various psychoacous��c metrics exist that are derived from sounds, e.g., Loudness, Sharpness, Roughness, etc. The
influence of these metrics on pleasantness of sound has been well documented. An ini��al aim would be to inves��gate correla��ons between EEG features that are
affected by these psychoacous��c metrics and/or that carry informa��on about the emo��onal state of the listener, based on parallels with studies on EEG‐based emo��on
affected by these psychoacous��c metrics and/or that carry informa��on about the emo��onal state of the listener, based on parallels with studies on EEG‐based emo��on
detec��on. The study will require EEG experiments in parallel with subjec��ve mock jury tests in which subjects listen to certain sounds of interest. Principles of machine
learning will be used to find pa��erns in the EEG which relate to emo��onal responses to sound.

Promoter: Prof. Alexander Bertrand (ESAT) ([email protected])
(Feel free to stop by for more informaᣆ�on)

Daily supervision: Abhijith Mundanad Narayanan (ESAT) ([email protected]) + regular interac��on with Siemens Leuven.

Work load: Literature 20% Experiments 30% Implementa��on 30% Valida��on 20%

1Website: http://www.plm.automation.siemens.com/en_us/products/lms/
2LMS Test.Lab Jury Testing: http://www.plm.automation.siemens.com/en_us/products/lms/testing/testlab/acoustic/jurytesting.shtml
EEGbased Auditory Attention Detection with Deep Neural Networks for Neurosteered Hearing
Prostheses – #36
by Das Neetha – last modified Mar 28, 2017 16:54
Promotors: Alexander Bertrand , Tom Francart (Dept. Neurosciences, KU Leuven) – Mentors: Neetha Das , Amir Hossein Ansari
Background:
Hearing impaired people often have difficulties to understand speech in noisy environments, which is why hearing aids are equipped with noise reduction
algorithms. However, in a multispeaker scenario, the noise reduction algorithm has difficulties to find out which speaker the subject is attending to, and
which speakers should be treated as noise. Recently, it has been demonstrated that the attended speaker can be detected based on
electroencephalography (EEG). This paves the way towards neurosteered hearing devices (Figure 1) that use auditory attention detection (AAD) to
steer noisesuppression algorithms towards the attended speaker.
There exists different approaches to perform AAD such as, e.g., solving a classification problem based on features extracted from EEG, or training a
spatiotemporal decoder to reconstruct the envelope of the attended stream from EEG. However, one problem that current AAD algorithms struggle with
is the relatively large data windows needed to detect attention, which affects the detection speed of these algorithms. Therefore we are looking for a
new datadriven method that can maintain high AAD accuracies, along with the ability to track attention at a faster rate.
Deep neural networks (DNNs) (Figure 2) have recently become very popular for, e.g., image processing, speech recognition, and other machine learning
problems, and have shown impressive results in various areas. In particular their combined feature extraction and classification capabilities are a major
strength of these approaches. In this thesis, we will investigate whether DNNs allow to perform EEGbased AAD, and whether they can improve
accuracy and detection times. We will focus on socalled convolutional neural networks (CNNs), which are currently the stateoftheart and the most
common technique in imageprocessing.
Goal and Objectives:
We currently work with trained spatiotemporal decoders to detect auditory attention from EEG recordings. To take a step further towards the design of a
better suited, faster and more reliable AAD algorithm, the feasibility of using CNNs for attention decoding will be investigated. The goal is to design a
deep CNN architecture that automatically extracts the relevant features from raw EEG+audio data and classifies which (of two) is the attended speaker.
We will benchmark the results with existing AAD techniques, to assess the added value of DNNs (in terms of feature extraction and/or classification
performance).

Required Skills:
Matlab experience
(Biomedical) signal processing
Patternrecognition
Recommendations:
We have already a few Matlab toolboxes for deep learning. But there are many powerful Python toolboxes implemented by Google, Microsoft and
other famous research teams in this field. If you know Python, this can be very useful (although not required, we can also work with Matlab
toolboxes).
Towards an improved phenotyping of obstructive sleep apnea – #39
by Varon Perez Carolina – last modified Mar 28, 2017 22:36
Promotors: Sabine Van Huffel , Carolina Varon Perez , Dries Testelmans (UZ Leuven) – Mentor: Margot Deviaene
Background
Sleep apnea is a sleeprelated breathing disorder characterized by repetitive cessations of breathing or airflow reductions during the night. During these
events, a typical behaviour on the electrocardiogram (ECG) is observed: bradycardia followed by tachycardia. In addition, these interruptions in the
“normal” breathing cycle cause reductions in the blood oxygen saturation (SpO2). Sleep apnea affects about 10% of middle aged adults, and it is
considered a risk factor for morbidity and mortality due to its longterm effect on the cardiovascular system. This effect is related to different
physiological mechanisms like systemic hypertension and increased sympathetic modulation that in a long term compromises the wellfunctioning of the
heart.
Currently, sleep apnea is diagnosed using polysomnography (PSG), which is a sleep test typically performed in a hospital, and where multiple signals
such as the ECG, the respiratory effort, and the SpO2 are recorded. From these signals and all other variables recorded during the PSG, different
parameters can be extracted to assess the severity of the disease. The most common parameter is the apnea/hypopnea index (AHI).
Even though the PSG and the AHI are very important diagnostic tools in sleep medicine, they have, on the one hand, limited predictive value of
cardiovascular risk and, on the other hand, they correlate poorly with sleep apnea therapy outcome. Therefore, it is clear that a better characterization of
apnea patients is needed, which will improve the understanding of the underlying pathophysiology of sleep apnea. Furthermore, a new characterization
can also help prioritize patients for treatment, ultimately reducing their morbidity and mortality.
Work Description
The main goal of this master thesis is to analyze the ECG, respiratory effort, and SpO2 signals recorded from patients suffering from sleep apnea, who
were referred to the sleep laboratory of the University Hospitals Leuven (UZ Leuven). These patients are grouped into two categories depending on how
they react to episodes of apnea. For instance, some patients will have more frequent arousals after an apnea, and other patients will have a more
reduced SpO2 level. These two types of events are typically observed in the signals recorded during the PSG, hence, it is the goal of this project to
develop algorithms that can detect automatically these different reactions. For example, different characteristics of the signals after episodes of sleep
apnea will be analyzed using time and frequency parameters. At this point, the student will need to extract informative features from the different signals
and analyze how they relate to the symptoms and the severity of sleep apnea.
This master thesis will be performed in close collaboration with the sleep clinicians at the sleep laboratory of the UZ Leuven. Candidates should have
basic knowledge of biomedical signal processing and programming skills in Matlab.
Multimodal detection of sleep apnea – #40
by Deviaene Margot – last modified Mar 29, 2017 10:31
Promotors: Sabine Van Huffel , Carolina Varon Perez , Dries Testelmans – Mentor: Margot Deviaene
Background
The obstructive sleep apnea syndrome (OSAS) is the most common sleeprelated breathing disorder, it affects around 2% of the middle aged women
and 4% of middle aged men. It presents itself as complete or partial cessations of breathing. These events can be caused by an obstruction in the upper
airway or by a loss of respiratory drive. Sleep apnea causes excessive sleepiness, poor concentration and an increased cardiovascular risk on the
longer term.
Diagnosis of sleep apnea is currently based on a full polysomnography (PSG) in the hospital. Manual scoring of the recordings is done. This is a very
costly and labor intensive process which is uncomfortable for the patient. Due to these drawbacks a long waiting list for PSG studies exists and some
patients are not eager to perform the study. As a result, up to 90% of patients with OSAS remain undiagnosed. Therefore, a lot of attention is on the
automation of the scoring process for screening sleep apnea, and on finding alternative measurements that can be performed in a home environment.
Work Description
The goal of this master thesis is to develop a multimodal algorithm to automatically detect the individual apnea events based on respiration, blood
oxygen concentration (SpO2) and electrocardiogram (ECG). The apnea event will cause a cessation of breathing, which can be detected in the
respiratory belt signals. This loss of respiration will result in a drop in SpO2 and/or an autonomic arousal (e.g. increase in heart rate). In order to detect
these events, the student should preprocess all signals, extract interesting features from each modality and combine these features in a multimodal
classifier. An algorithm for detection of apneas based on ECG is already available and can be used, so that the student can focus on the processing of
the respiratory and SpO2 signals. PSG data containing those signals will be available from the Sleep Monitoring Unit at UZ Leuven, the apnea events
are scored by sleep experts.
This master thesis will be performed in close collaboration with the sleep clinicians at the sleep laboratory of the UZ Leuven. Candidates should have
basic knowledge of biomedical signal processing and programming skills in Matlab.
Epileptic source localization in presurgical planning – #64
by Hunyadi Borbála – last modified Mar 30, 2017 13:27
Promotors: Sabine Van Huffel , Wim Van Paesschen , Borbála Hunyadi – Mentor: Ying Gu
Background
Epilepsy is a neurological disorder, affecting 1% of the worldwide population. The manifestation of this disease is the epileptic seizure, an abnormal,
synchronous electrical activity of the neurons in the brain. More than 30% of epilepsy patients continue to have seizures despite of medication, hence
their quality of life is seriously compromised. Surgical resection of the epileptogenic focus might offer cure for these patients. The success of the
surgery strongly depends on the accurate identification of the brain region responsible for the generation of seizures, i.e. the epileptogenic zone.
As epileptic seizures ictal activity and epileptic spikes interictal epileptic activity are electrical phenomena, they can be observed by
electroencephalography (EEG), a noninvasive recording technique which measures the electrical potentials generated in the brain. Although the
temporal resolution of EEG is very good, the spatial resolution is limited by the number of electrodes and the fact that the electrodes are placed on
scalp. EEG source imaging (ESI) aims at localizing the active areas in the whole volume of the brain based on the small set of electrodes placed on the
surface of the head. As such, ESI is a very challenging engineering problem.
A complementary noninvasive technique, functional magnetic resonance imaging (fMRI) measures the socalled bloodoxygenleveldependent (BOLD)
signal. As active brain tissue consumes oxygen, the BOLD fMRI signal fluctuates in response to brain activity, i.e. in response to epileptic activity.
fMRI is commonly analyzed using EEG information, where the timing of epileptic activity is used to construct a model time course for the fMRI. The
advantage of fMRI is that it provides information with a very good spatial resolution. However, its accuracy depends on the goodness of the EEGbased
model.
Content
The aim of this thesis is to apply and evaluate novel ESI techniques on real clinical data. The EEG data are recorded in the epilepsy monitoring unit of
UZ Leuven as part of the standard presurgical evaluation. The results of ESI should be compared to the surgical resection zone as gold standard in
patients where a successful operation was performed. Moreover, they should also be compared to the outcome of EEGcorrelated fMRI analysis, as an
alternative diagnostic tool. ESI should be performed on EEG data inside and outside the fMRI scanner. It should be investigated whether the available
ESI tools are robust enough against scanner artifacts. Possibly, advanced preprocessing tools should be applied which can separate the activity of
interest from these artifacts. Finally, if ESI is successful inside the scanner, ESI of EEG and EEGfMRI can be integrated into a single, more sensitive
multimodal analysis tool.
Multimodal artefact removal in ECG signals – #63
by Moeyersons Jonathan – last modified Mar 30, 2017 13:59
Promotors: Sabine Van Huffel , Carolina Varon Perez – Mentor: Jonathan Moeyersons
The electrocardiogram (ECG) is a fundamental part of screening programs for the detection of heart disease. It reflects differences in transmembrane
voltages in myocardial cells that occur during depolarization and repolarization within each cardiac cycle. Ideally these differences can be recorded from
the heart itself, but this is, due to obvious reasons, not very practical. Therefore the differences are recorded by electrodes placed on the skin. Inevitably
this comprises measurement of variations in skin impedance and muscle contractions as well.
The aim of this master thesis is to develop an algorithm capable of detecting artefacts in ambulatory ECG signals. Artefacts are electrical signals that
are measured by the ECGelectrodes, but do not originate from the heart. Since they can have large amplitudes, they introduce erroneous results and
cause misdiagnosis or inappropriate treatment decisions. Motion artefacts represent the most challenging source of noise in ambulatory ECG signals,
because they share the same frequency spectrum as the ECG signal. Furthermore, the shape of motion artefacts can be very similar to ECG
waveforms.
A recent method, developed by the STADIUS research group characterizes ECG segments by means of the AutoCorrelation Function (ACF) [1]. Since
the ACF shows repetitive patterns, and therefore also disturbances, such as artefacts, it can be used to distinguish between contaminated and clean
segments. The advantage of this method is that a level of noise is assigned to each ECG segment, making it possible to identify and locate the
contaminated ones. However, it remains very difficult to accurately detect artifacts using only the ECG signal.

The research will be performed in close collaboration with imec, who provide a dataset of 105 subjects. The measuring device (Stingray, imec) used to
measure the data contains a twochannel ECG and a triaxial accelerometer. The data were recorded in view of the imec SWEET study, to be used for
stress analysis. The challenge in this thesis is to optimally remove the artefacts from the ECG , e.g., by exploiting the information in the accelerometer
signals and integrating it in the existing algorithm or by developing a novel multimodal algorithm. One example is developed by Tong et al. [2].
The result of this master thesis is an automated, offline and robust algorithm, capable to detect and locate artefacts, based on multimodal signals in
ambulatory ECG.
The student should have experience in writing Matlab code and should have followed the course “Biomedical Data Processing (BKULH03I2A)”.
[1] C. Varon, D. Testelmans, B. Buyse, J. A. K. Suykens, and S. Van Huffel, “Robust artefact detection in longterm ECG recordings based on
autocorrelation function similarity and percentile analysis,” Proc. Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. EMBS, pp. 3151–3154, 2012.
[2] D. Tong, K. Bartels, and K. Honeyager, “Adaptive reduction of motion artifact in the electrocardiogram,” EMBS/BMES Conference, Proc. of the
Second Joint, vol. 2, pp. 1403– 1404, 2002.
Epileptic seizure detection with wearable EEG and ECG: feature integration – #2
by Vandecasteele Kaat – last modified Mar 30, 2017 12:27
Promotors: Sabine Van Huffel , Borbála Hunyadi , Wim Van Paesschen (UZ Leuven) – Mentors: Ying Gu , Kaat Vandecasteele
Background
Approximately 1% of the world population exhibits symptoms of epilepsy, a serious disorder of the central nervous system. People with epilepsy suffer
from recurrent seizures that occur at unpredictable times and usually without warning. Frequent seizures increase the risk of sustaining physical
injuries. There is a need for quick automatic seizure onset detection and afterwards intervention during daily life. Currently, clinical monitoring of seizure
is performed visually by clinicians. The use of an automatic seizure detection would drastically decrease the workload of clinicians. Furthermore it will
facilitate the followup of the disease evolution and the response to medication.
This thesis proposal is part of an ICON project, aiming to develop a wearable seizure detection device. Its development involves an interdisciplinary
team: Byteflies for wearable solutions, Pilipili for product development design, UCB for bringing medical quality products to the market and UZ
Gasthuisberg for patient monitoring. KU Leuven’s responsibility is algorithm development for seizure detection. In order to improve the accuracy of
seizure detection, multiple physiological signals are considered: here EEG and ECG. For the purpose of a comfortable home application, wearable
sensor systems are selected.
Work description
The EEG measures the electrical activity of the brain and is a wellestablished technique in epilepsy diagnosis and monitoring. The epileptiform EEG is
characterized by its spectral, temporal and spatial distribution. The relevant features are extracted by various methods, including FFT, autoregressive
modeling, wavelet transform, phase synchronization, entropy, spatial filtering and so on. In addition to the EEG, ECG is affected during a seizure. From
the ECG signal, the tachogram can be calculated. Features like heart rate increase, maximal heart rate are extracted. Machine learning methods (for
example LSSVM) will be applied based on EEG and ECG features together. The thesis project will focus on EEG and ECG optimal feature selection
and integration. The performance of seizure onset detection can be compared between different feature integration strategies and between different
physiological signals.
Basic Matlab programming and basic knowledge about signal processing are required. The student should be interested in machine learning techniques.
Autonomic Nervous system maturation in preterm babies based on Heart rate variability – #70
by Lavanga Mario – last modified Mar 30, 2017 14:27
Promotors: Sabine Van Huffel , Gunnar Naulaers , Alexander Caicedo Dorado – Mentors: Mario Lavanga , Ofelie De Wel
TITLE
Autonomic Nervous system maturation in preterm babies based on Heart rate variability
GUIDANCE
For more information, please contact: Mario Lavanga
([email protected])
Promotors: Sabine Van Huffel
Copromotor:
Gunnar Naulaers, Alexander Caicedo Dorado
Supervisor: Mario Lavanga, Ofelie De Wel
CONTEXT (e.g. Information about the problem definition, research project of which the master’s thesis
is part)
According to the European Task force on Heart Rate Variability (HRV), the instantaneous heart rate
fluctuations reflect the control of Autonomic Nervous System (ANS) on the cardiac muscles. The long
term variations (known as lowfrequency, LF) represent, mainly, the sympathetic stimulation, while the
shortterm oscillations (known as high frequency, HF) are, mainly, induced by the parasympathetic or
vagal system. Although the autonomic nervous system is present since birth, the autonomic control is
subject to the individual development and is known to be immature in preterm babies. In particular, the
vagal tone undergoes the major development.
HRV oscillations can be investigated by analysis of the tachogram, which represent the variation in time
between 2 consecutive heart beats.
Spectral analysis is one of the most common tools used to extract the power contributions of LF and
HF, although there are different approaches (both in frequency and timedomain) to compute the same
contributions.
Recently it has been shown that respiration, EEG and HRV patterns change according to the different
sleep stages. Therefore, research efforts have been focused on investigating how sleep stages affect
the coupling among these signals. To the best of our knowledge, there are no studies, yet, that show
how this coupling evolves according to the newborn development.
GOAL
The aim of this thesis is to describe the maturation of autonomic nervous system based on HRV
analysis on a database of preterm babies with normal development. The candidate will be asked to
compute different linear and nonlinear features, from the tachogram, and describe their evolution at
different ages using different regression methods. Furthermore, the candidate will be asked to
investigate the evolution of the cardiorespiratory and cardiocerebral couplings, which will be estimated
using information dynamics methodologies, among others.
RESULT
The final results from this thesis should produce a maturation chart for the autonomic nervous system,
as well as for the cardiorespiratory and cardiocerebral coupling. These charts should be presented in
such a way that they are easy to interpret by clinicians.
PROFILE (e.g. rather theoretical / rather practical implementation, foreknowledge (courses, methods,
computer language(s) et cetera))
The student should have some experience in writing Matlab code and should have followed the course
“Biomedical Data Processing (BKULH03I2A)”.
1 student 1 student
3. Biomedical Imaging
Segmentation of joints in CT images – #14
by Delaere Dominique – last modified Mar 28, 2017 09:33
Discipline: Biomedical Imaging
Promotor: Dirk Vandermeulen – Mentor: Pieter Van Leemput (Medicim)
Background
Medical images, like Computed Tomography (CT) images, are typically noisy and contain artifacts. As a result, separate objects can appear connected
in the image. A 3D surface model generated from such image data will only contain one part that represents both objects. This prevents the different
objects from moving relative to each other in the scene.
Various objects in the image are typically separated from each other using an image segmentation method. A vast number of segmentation algorithms
exist. Most of them search for the boundaries of one of the objects. In many cases, the user has to manually correct the end result due to the
connecting (CT) artifacts.
Content
At Nobel Biocare – Medicim, we have developed a novel segmentation algorithm to segment objects or parts of objects in the image scene, especially
in the case when they appear to be connected through (CT) artifacts. An example of our segmentation of a CT image of the hip joint is shown in the
figure above. The left figure shows a surface representation of a hip joint where the femoral head and the pelvis are connected to each other through CT
artifacts. The right figure illustrates the separation of the femoral head from the pelvis by the algorithm (note the different colouring).
In this thesis, we want to investigate how the algorithm can be applied to other – mostly orthopaedic – applications, e.g. segmentation of other types of
joints, fractures or the bones in the hand. These applications will require modifications and additions to the existing algorithm as well as a thorough
validation study.
Practical
The candidate is expected to develop the required software either in C++ or Matlab. Basic programming skills in either one of these programming
languages are a necessity.
Location
The preferred work location is the Nobel Biocare – Medicim office, Stationsstraat 102 in Mechelen (next to the central train station). We expect the
interested candidate to work on his master thesis for the complete academic year (September / October to May).
Automated segmentation of the mandibular nerve canal from CBCT images – #18
Promotor: Dirk Vandermeulen – Mentors: Bart Goris (Medicim) , Johannes Keustermans (Medicim)
Background
The prosthetic restoration of missing or corrupted teeth greatly affects the quality of life because teeth have a large influence on one’s social interactions
and wellbeing. This is due to the important functions of teeth such as food digestion, speech, and maintaining the shape of the lower face. A routinely
used, reliable, and well established technique for prosthetic restoration of teeth is based on dental implants. Dental implant based treatments target a
functional and aesthetic outcome with minimal discomfort for the patient and short treatment times. A careful and detailed planning of the dental
implants, as well as adequate tools to transfer the planning to the operating room are key factors for the success of the entire treatment. The accurate
assessment of the patient’s anatomy, often based on 3D medical images, is a prerequisite for such a planning. Thereby it is of utmost importance to
prevent damage to the facial nerves, as the latter could lead to severe discomfort, pain, inflammation or even facial analgesia. The mandibular nerve is
in this aspect the most critical, as illustrated in the figure below.
Objectives
The objective of this thesis is to develop a solution to automatically visualize the mandibular nerve canal in the 3D dental implant planning software. The
nerve canal can be detected based on the surrounding image intensities and its specific shape. Different algorithms will be investigated to find the
location of the nerve canal. Using a visual aid of the nerve as in Figure 1, the position of the dental implants can be judged accurately to avoid contact
with the nerve. As such, damage to the mandibular nerve can effectively be prevented. In order to visualize the mandibular nerve canal, it should be
detected in the 3D (CB)CT images.
Work description
3D CBCT images are abundantly available. Our preliminary research yielded a first tentative concept for solving the proposed problem. As part of the
thesis this tentative concept can be further elaborated and needs to be compared to alternative options available from the relevant literature. The
resulting segmentation(s) can be compared to a ground truth manual segmentation of the nerve canal obtained with our software.
Practical
Location
Visualizing subvoxel structures in 3D surface renderings of (CB)CT images – #19
Promotor: Dirk Vandermeulen – Mentor: Bart Goris (Medicim)
Background
Cone beam computed tomography is nowadays a common technique to obtain 3D structural information of the human anatomy. A proper visualization of
the dataset is crucial to obtain accurate diagnostic conclusions from the corresponding 3D (CB)CT image. In many applications, a 3D visualization is
obtained by constructing a 3D surface that connects all voxels having similar intensity levels. An example of such a 3D isosurface rendering is
visualized in Figure 1. However, it is commonly known that a single isovalue cannot represent anatomical structures at subvoxel level. This is due to
the partial volume effect, which mixes the gray values of nearby tissues into a single voxel. This leads to lower gray levels for thin skull features. The
single global isovalue fails to capture these features yielding holes in the final surface representation of the skull. Two examples are indicated by the red
arrows in Figure 1.
Objectives
In this work, the student will optimize stateoftheart visualization techniques so that the isovalue that is used for the visualization can be adapted
locally. This can either be achieved by calculating the isovalue over smaller volumes of interest or by incorporating other types of information, like local
image gradients or curvature that better represent the local thin structures. Consequently, anatomical structures at subvoxel level can be made visible
and partial volume artefacts can be counteracted. This yields improved visualizations from the 3D (CB)CT scan. Ultimately, this may lead to better
diagnostics and better patient care once the new visualization technique is implemented in medical software.
Work description
3D CBCT images are available together with the 3D visualization software. First, a review of the relevant visualization algorithms should be performed.
Next, based on the literature review, a dedicated algorithm needs to be implemented that adaptively determines the local isovalues used for the 3D
visualization.
Practical
Location
Exploring the heritability of facial features using spatiallydense morphometrics – #27
Promotor: Peter Claes – Mentor: Hanne Hoskens
Background
The face is a biological billboard of our identity and underlying genes. Indeed, there is elaborate evidence that facial features are under strong genetic
control. Such evidence includes remarkable facial similarity between identical twins and clear facial resemblances within families.
Studies using twin and parentoffspring databases can separate environmental from genetic factors on facial morphology which is of great interest to
future genetic facial shape investigations and applications. Most if not all studies on craniofacial heritability start from sparse descriptions of facial
shape, e.g. by measuring geometric features such as distances and/or angles. However, these representations typically overlook salient features of
facial shape. Furthermore, heritability studies today do not investigate coinheritance between different geometric features.
The aim of this thesis is to perform a spatiallydense analysis of facial heritability and coheritability.
Content
The student will make use of a database of 3D facial surface scans of parents and offspring. In a first instance, a spatiallydense view on facial
heritability per landmark will be provided. In a second instance, a subdivision of facial shape into modules of coinheritance will be provided.
Subsequently, the benefit of such a modularization will be tested in a limited association study between facial shape and known craniofacial genes.
Practical
Basic programming skills (MATLAB)
Location
ESAT/PSI, Medical Imaging Research Center, UZ Gasthuisberg
Deep convolutional neural networks for automated segmentation of organsatrisk in longitudinal
CT scans for proton therapy – #25
Promotor: Frederik Maes – Mentors: David Robben , Wouter Crijns (Radiotherapy)
Background
Radiotherapy aims at eradicating a tumor through accurate delivery of ionizing radiation. Before therapy can start, a patient specific planning is made as
to deliver the required dose to the tumor and keep the radiation of other organs below specified thresholds. Hereto, a CTscan of the patient is acquired
and all the organs at risk are delineated, as illustrated in Fig. 1.a. Manual delineation of these organs is a timeconsuming and tedious task which takes
up to 2 hours for the headneck region.
In the near future, UZ Leuven will offer proton therapy, where the dose is delivered via protons instead of photons. Proton therapy offers more accurate
dose delivery, but requires more accurate planning – up to the point that the shrinkage of the tumor and surrounding tissue is problematic. This will be
countered by acquiring a new CT scan before the start of each radiation session, and updating the segmentations of the tumor and organsatrisk as
well as the dose planning. To keep this approach feasible from a practical point of view, this process needs to be automated. In this thesis, we aim to
develop robust image analysis methods to automate the segmentation of the organsatrisk in a CT scan, taken into account the manual delineations of
previous scans.
Content
We recently developed a proofofconcept that demonstrates that deep learning

segmentation techniques, in particular convolutional neural networks, might be able to What is deep learning?
delineate the organsatrisk in a CT scan. However, in the context of repeated
scanning, the manually corrected (and hence ground truth) delineation of the In traditional learning methods, a classifier learns to
previously acquired CT scan can provide great addition guidance for the next scan distinguish voxels of different organs based on a set of
and should not be discarded. You will investigate deep learning techniques that allow precalculated features such as the local average, the
to delineate the organsatrisk in a CT scan – as learned from an annotated training Laplacian, the Hessian, …
set – while taking into account an earlier acquired and delineated scan of the same
patient. This approach has been applied successfully to several
applications, but suffers from the inherent drawback that
In this thesis, you will first become acquainted with deep learning techniques and in features are application specific and hence intensive
particular with convolutional neural networks. Subsequently, you will explore novel experimentation is required to find the best ones. In deep
techniques to address the earlier mentioned challenge. We provide a large database learning methods, the classifier learns to distinguish voxels
of manually delineated CTscans, the required computational infrastructure (powerful from different organs based on the raw image, by
graphic cards), the working proofofconcept and supervision with both technical and simultaneously learning the best features and how to use
clinical expertise. them. In recent benchmarks, deep learning techniques
systematically outperform the traditional methods.
Prerequisites:
good mathematical skills
strong interest in machine learning techniques
able to or willing to learn to program in Python
Location
Deep convolutional neural networks for automated segmentation of organsatrisk in radiotherapy –
#26
Promotor: Frederik Maes – Mentors: David Robben , Wouter Crijns (Radiotherapy)
Background
Radiotherapy aims at eradicating a tumor through accurate delivery of ionizing radiation. Before therapy can start, a patient specific planning is made as
to deliver the required dose to the tumor and keep the radiation of other organs below specified thresholds. Hereto, a CTscan of the patient is acquired
and all the organs at risk are delineated, as illustrated in Fig. 1. Manual delineation of these organs is a timeconsuming and tedious task which takes up
to 2 hours for the headneck region. In this thesis, we aim to develop robust image analysis methods to automate the segmentation of these organs.
Content
We recently developed a proofofconcept that demonstrates that deep learning
segmentation techniques, in particular convolutional neural networks, might be What is deep learning?
able to automate this task. You will further improve this proveofconcept,
aiming at obtaining a fully functional method. Along the way, several important In traditional learning methods, a classifier learns to distinguish
research and design questions need to be addressed such as what is the voxels of different organs based on a set of precalculated
optimal architecture of the neural network, how can the training set be features such as the local average, the Laplacian, the Hessian,
artificially enhanced or how should we perform automatic quality control of the …
resulting segmentations.
This approach has been applied successfully to several
In this thesis, you will first become acquainted with deep learning techniques applications, but suffers from the inherent drawback that features
and in particular with convolutional neural networks. Subsequently, you will are application specific and hence intensive experimentation is
explore novel techniques to address the earlier mentioned challenges. We required to find the best ones. In deep learning methods, the
provide a large database of manually delineated CTscans, the required classifier learns to distinguish voxels from different organs based
computational infrastructure (powerful graphic cards), the working proofof on the raw image, by simultaneously learning the best features
concept and supervision with both technical and clinical expertise. and how to use them. In recent benchmarks, deep learning
techniques systematically outperform the traditional methods.
Prerequisites:
good mathematical skills
strong interest in machine learning techniques
able to or willing to learn to program in Python
Location
Multi Resolution Decomposition for Face Recognition – #29
Promotor: Peter Claes – Mentor: Dzemila Sero
Background
Face biometrics studies the facial shape in order to build a recognition system able to identify an unknown individual (for example, on a crime scene) or
verify whether a user should be given access, for instance, to a specific building. The standard characteristics used as identifiers are the face,
fingerprints, iris and voice. At the Medical Imaging Research Center, we build recognition systems such that information like sex, age, height, weight,
ancestry is predicted from the face and then compared to a target unknown face, for which the correspondent string information is extrapolated (in real
case scenarios, these could be from a blood or saliva sample, for example). The combination of multiple matching scores gives the likeliness of the
target face to match each of the faces present in the database. Given the central role of the face in determining a successful classification task, its
representation needs specific care. Multiple facial characteristics have been investigated: principal components, radial and stream lines, linear distances
between specific points on the face, and so on. The goal of the current thesis is to implement a wavelet decomposition of the face, and use its
coefficients as predictors. Another key topic is the combination of scores. Literature on combination of multiple scores proposes three methods: the sum
rule, the classification based fusion, and density based score fusion. The second main goal of the thesis is implementing a density based score fusion.
Keywords – multiresolution, hierarchical modeling, statistical shape modeling, facial recognition, biometric score level fusion
Content
The student will be asked to:
implement a wavelet decomposition on the 3D facial meshes;
implement a classifier able to predict soft traits from facial wavelets;
implement a biometric recognition system wherein a good matching score fusion technique should be included;
evaluate the results with the standard approaches such as the Identification (Cumulative Matching Curves) and Verification curves (Equal Error
Rate value).
(optional): combine the wavelet decomposition with the inhouse facial representation methodology
Practical
Basic programming skills (preferably Matlab)
Workload
20% literature & gaining experience
50% development & implementation
20% interpretation & validation
10% dissertation & presentation
Location
Convolutional Neural Network for face recognition – #33
Promotor: Peter Claes – Mentor: Dzemila Sero
Background
Face biometrics studies the facial shape in order to build a recognition system able to identify an unknown individual (for example, on a crime scene) or
verify whether a user should be given access, for instance, to a specific building. The standard characteristics used as identifiers are the face,
fingerprints, iris and voice. At the Medical Imaging Research Center, we build recognition systems such that information like sex, age, height, weight,
ancestry is predicted from the face and then compared to a target unknown face, for which the correspondent string information is extrapolated (in real
case scenarios, these could be from a blood or saliva sample, for example). The combination of multiple matching scores gives the likeliness of the
target face to match each of the faces present in the database. Given the central role of the face in determining a successful classification task, its
representation needs specific care. Multiple facial characteristics have been investigated: principal components, radial and stream lines, linear distances
between specific points on the face, and so on.
The goal of the current thesis is to implement a convolutional neural network able to construct invariant descriptors of the face, and use its coefficients
as predictors. The second goal, is to use these facial descriptors in a biometric system.
Keywords Deep Learning, Convolutional Neural Networks, Shape Descriptors, Facial Recognition, Biometrics
Content
The student will be asked to:
implement a convolutional neural network from the 3D facial meshes;
implement a biometric recognition system;
evaluate the results with the standard approaches such as the Identification (Cumulative Rank Curves) and Verification curves (Equal Error Rate
value).
Practical
Basic programming skills (preferably in Matlab)
Workload
Location
Ancestral Faces – #32
Promotor: Peter Claes – Mentor: Jiarui Li
Background
DNA extracted from a blood sample or a strand of hair collected at a crime scene is useful for identifying criminals. Nowadays, it is even possible to
reconstruct the facåe of the criminals just from DNA [1], in absence of an eyewitness. Many parts of the face are affected by ancestry, which can be
estimated from DNA as well. Therefore, ancestral faces, constructed from estimated ancestry axes, reflect the major facial variation determined by the
ancestry information (e.g. see Figure 1). The ability to reconstruct someone’s ancestral face is important in the final prediction of faces from DNA.
Content
Practically, DNA found at forensic crime scenes, is typically processed with a variety of DNA typing platforms. The result is that the actual information
extracted from one platform is different compared to the information extracted from another platform. However, our current construction of ancestral
faces requires a fix DNA information input. Therefore, the aim of this project is to provide an interface for the construction of ancestral faces, such that
this technique is applicable to any random set of DNA input information. This involves an online relearning of the prediction of ancestry faces in
function of the DNA information given. The student can start with the current implementation and expertise in the lab.
AIM
Given the reference dataset (1000 GENOME Project [2]) with ancestry labels, designing algorithms and software which can:
Given individuals with any subset of DNA information of the reference dataset, estimate the comparable continuous axes which represent the
ancestry information
Generate ancestral faces given the estimated ancestry axes
References
Claes P, Liberton D K, Daniels K, et al. Modeling 3D facial shape from DNA[J]. PLoS Genet, 2014, 10(3): e1004224.
1000 Genomes Project Consortium. A global reference for human genetic variation[J]. Nature, 2015, 526(7571): 6874.
Practical
Programming Languages: MATLAB and/or C/C++
Workload
Location
Deep learning on 3Dfacial images to identify phenotypic differences between two groups – #28
Promotor: Peter Claes – Mentor: Karlijne Indencleef
Background
The face is the most identitycoding part of the human body. Allthough every face is unique, facial characteristics are influenced in a specific way by
sex, age, BMI, ethnicity and other genetic or environmental factors. Using 3D facial images, this delivers very interesting research objectives.
One of these objectives is part of unraveling the genetic archtecture of nonsyndromic cleft lip with or without cleft palate (NSCL/P). Because first
degree relatives of NSCL/P patients have a high chance of carrying susceptibility genes for the condition, it can be useful to focus on their phenotype.
Our research unit has confirmed the presence of a facial endophenotype of NSCL/P with advanced 3D morphometric techniques using a spatiallydense
network of quasilandmarks.
The phenotypic differences between NSCL/P relatives and controls are subtle. This is why the method of analysis of facial phenotypic differences
between two groups will be developed on sex differences. Currently, these differences are determined by partial least square regression models and the
variables of interest are predefined. Deep learning offers new possibilities to identify facial features that are chaacteristic for the variable of interest.
Content
The goal of this thesis is to develop and apply deep learning algorithms on a database of 3D facial images to identify distinctive facial features for
women and men. The probability for a 3D–image to be in the male or female group will offer a continuous score in terms of feminity and masculinity in
the face. This poses two challenges:
Developing a deeplearning algorithm developed for 3Dfacial images
Applying the deeplearning algorithm on a dataset that is less extensive than applied in usual deep learning analyses
Location
Alzheimer prognosis using imagebased shape features – #41
Promotors: Paul Suetens , Peter Claes – Mentor: Dorothy Gors
Background
Alzheimer’s Disease (AD) causes severe memory loss and cognitive decline. When imaging the brain using magnetic resonance imaging (MRI), it is
observed that AD also changes the shape of the brain, which makes shapebased features extracted from MRI a suitable biomarker for AD. Less
known than AD is mild cognitive impairment (MCI), where patients have only moderate memory and cognitive deteriorations. However, for most of these
patients their illness progresses towards AD (MCIc, MCI converters). Only a few MCI patients remain stable (MCInc, MCI nonconvertors), and it is
utterly important for the patient’s quality of life to predict or monitor whether he or she will or is converting to AD or not? A possible prognosis could be
made by comparing brain shape of a MCI patient to healthy controls and AD patients. The underlying hypothesis would be that the resemblance in brain
shape of MCI patients to AD patients indicates the progression of the memory loss and cognitive decline.
Content
There are two kinds of voxelbased variables to describe brain shape from MRIscans: the voxelwise tissue densities (grey matter (GM), white matter
(WM) and cerebrospinal fluid (CSF) densities) or the voxelwise volume differences (Jacobians). For this thesis project the student will first implement a
known pattern recognizer referred to as COMPARE, developed by Fan et al.[1], to make the MCI conversion predictions based on GM densities.
Secondly, the student will evaluate COMPARE against an alternative inhouse pattern recognizer using Jacobian information. Finally, the student is
encouraged to experiment with the combination of both techniques.
[1] Fan, Y., Shen, D., Gur, R.C., Gur, R.E., Davatzikos, C., 2007. COMPARE: Classification Of Morphological Patterns using Adaptive Regional
Elements. IEEE Trans. Med. Imaging 26, 93–105.
Practical
The student will have access to the tissue densities and Jacobians of about 600 MRI scans, together with inhouse matlabsoftware of a pattern
recognizer for AD prognosis.
Location
Additional Info:
Dorothy Gors will be replaced by Dzemila Sero on the thesis fair.
Motion correction for T2 mapping using cardiac MRI – #60
Promotor: Frederik Maes – Mentors: Sofie Tilborghs , Tom Dresselaers
Background
Cardiac magnetic resonance imaging (MRI) offers the ability to acquire anatomical and functional images of the heart in a noninvasive way with
sufficiently large spatial and temporal resolution. By changing the parameters of the MRI acquisition protocol, different images can be obtained during a
single imaging session that provide complementary diagnostic information about the morphology of the heart, the heart functioning and the cardiac
tissue itself, that can be quantified by suitable image analysis. Currently, image contrast changes are however visually (subjectively) interpreted. New
methods aim at providing quantitative parameter maps that allow a more objective evaluation (across patients / readers / centers) and may have added
value in case of nonfocal pathology (difficult visual interpretation). We previously developed an image analysis approach for T1 mapping of the heart,
which involves the quantification of the T1 relaxation time of the tissue from a series of images with different T1 weighting. However, the acquisition of
the images is complicated by the intrinsic motion of the heart and by breathing related motion which can be corrected by applying dedicated registration
algorithms. In this thesis, we wish to develop a similar processing strategy for T2(*) mapping of the heart.
Content
In myocardial T2 mapping a series of T2weighted images (‘echoes’) is acquired followed by a pixelwise exponential fit to derive the T2 in each pixel
(i.e. T2 map). This series of ‘echoimages’ is currently acquired segmented across multiple heart beats. Any motion will result in blurred images and
erroneous T2 fitting/mapping results. By altering the data acquisition (cfr. [1]) and applying appropriate motion registration algorithms in postprocessing,
motion can be corrected which should result in more reliable T2 estimates. The aim is therefore to evaluate the impact of this altered acquisition
strategy and offline motion correction (i.e. image registration) on T2(*) estimation. The final outcome would be a T2 mapping cardiac MRI method
that provides diagnostically useful information in patients with limited or absent breath hold compliance (about 50% of patients).
[1] Giri et al., Myocardial T2 mapping with respiratory navigator and automatic nonrigid motion correction, Magnetic Resonance in Medicine, 2012
Practical
Clinical data will be analyzed from patients that require T2 mapping MRI for diagnostics purposes. In case of breath hold noncompliance (blurred data),
a modified T2 mapping acquisition (freebreathing) will be added that provides data for offline motion correction and T2 fitting. Potentially study data
from healthy volunteers (e.g. athletes study) will also be included.
Prerequisites: This project will include data processing using different software packages (Matlab, ImageJ). Some experience with Matlab is beneficial
but not mandatory.
Collaborators: This project will involve, besides the promotor, the following people (of multidisciplinary background): Tom Dresselaers (physicist,
PhD), Sofie Tilborghs (ir., PhD student), prof. Jan Bogaert (radiologist, MD).
Location
Segmentation of magnetic resonance images of the spinal cord in neurodegenerative and
neuroinflammatory diseases – #56
Promotor: Frederik Maes – Mentors: Annemie Ribbens (icometrix) , Dirk Smeets (icometrix) , Diana Sima
Background
Neurodegenerative diseases are characterized by atrophy of brain tissue, but also atrophy of spinal cord tissue. For patients with multiple sclerosis
(MS), for instance, it has been shown that the severity of cervical spinal cord atrophy is correlated with the degree of disability and the severity of the
MS form [1]. The brain and spine of patients with MS or other neuroinflammatory diseases is also characterised by the presence of typical white matter
lesions. These can be visualized with Magnetic Resonance Imaging (MRI). Reliable measurements characterising the spine are becoming important for
monitoring disease progression in these patients.
[1] Rocca MA, Horsfield MA, Sala S, Copetti M, Valsasina P, Mesaros S, Martinelli V, Caputo D, StosicOpincal T, Drulovic J, Comi G, Filippi M. A
multicenter assessment of cervical cord atrophy among MS clinical phenotypes. Neurology. 2011;76:2096–2102.
[2] Images taken from http://www.radpod.org/2007/09/27/leukomyelitis/ (credit: Dr Frank Gaillard).
Content
icometrix has developed methods for lesion volumetry and atrophy estimation using brain MRI for clinical monitoring of patients with MS and other
neurological diseases. During the master thesis the students will research various diseaserelated spinal cord MRI biomarkers, and will study, compare
and adapt existing methods for spinal cord tissue and lesion segmentation, as well as for atrophy computation. The accuracy and reproducibility of the
evaluated methods will be estimated on collections of MRI datasets with available manual segmentation and on testretest MR images from MS
patients.
Prerequisites: 1 or 2 students. Each of the 2 students will focus on different biomarkers for lesion volumetry and atrophy estimation, respectively.
Workload
The thesis project consists of literature review of various diseaserelated spinal cord MRI biomarkers (20%), followed by implementation in
Python based on existing (opensource and inhouse) software tools (40%) and validation on MRI data (40%).
Location
The thesis project will be conducted in close collaboration and at the premises of icometrix, a spinoff company of KU Leuven (Kolonel Begaultlaan 1b /
12, 3012 Leuven, www.icometrix.com). icometrix specializes in providing services for quantification of brain MRI scans, in particular for diagnosis of
MS.
Semiautomatic spine segmentation for the creation of a subjectspecific musculoskeletal model in
an Adult Spinal Deformity population. – #61
by Scheys Lennart – last modified Mar 30, 2017 11:23
Promotors: Dirk Vandermeulen , Lennart Scheys , Ilse Jonkers – Mentor: Thomas Overbergh (KU Leuven (PhD Student))
Background: Spinal deformity is a type of pathology which consists of a complex interaction between structural deformities of the spine and
biomechanical compensations, heavily impacting on quality of life. As a result of this complexity, it is still unclear how spinal deformities influence the
dynamic biomechanics of a patient. The combination with the growing number of subjects suffering from this pathology makes it an emerging research
area.
Instrumented motion analysis offers the potential to assess the gross pre and post operative functionality in 3D. In combination with a mathematical
model representing the relevant musculoskeletal anatomy of a specific subject, i.e. a personalized musculoskeletal model (MSM, fig. 1 (right)),
biomechanical characteristics of the spine can be quantified in more detail. Today, this MSM is manually personalized to the patient’s anthropometry
based on stereoradiograpic images, acquired using a novel stateoftheart, lowdose biplanar imaging system (EOS Imaging, http://www.eos
imaging.com/) (fig. 2). This personalization is performed in a currently being developed software tool (fig. 3), which furthermore has the required import
and export capabilities to the software (OpenSim) typically used for further analysis.
Objectives: Within this master project we aim to develop a methodology to further increase the timeefficiency and accuracy of this critical
personalization step by defining a methodology to (semi)automatically extract the skeletal spinal alignment from biplanar radiographic images and
integrate this in the current software tool. Thus, we aim to answer the increased need for 3D personalized models in this pathology.
Work description: Given two orthogonal full body radiographic images, the main aim of this thesis project is to develop a method to
(semi)automatically accomodate the 3D position and shape of generic vertebrae untill they optimally match the radiographic projections visible on both
orthogonal images (fig. 4). Following integration in the already available software tool, we furthermore plan to document the validity, repeatability and
feasibility of this technique for extracting relative positions and orientations of neighbouring vertebrae.
Additional Info:
Special circumstances / required skills: The ideal candidate has a profound interest in medical image processing techniques and has experience with
(or the willingness to learn) programming.
Location where work will be performed: The working location is flexible. Daytoday work can be performed from home or the offices at the Institute
for Orthopeadic Research and Training (IORT, http://gbiomed.kuleuven.be/iort), located both at the University Hospitals Gasthuisberg (Leuven) and the
University Hospital Pellenberg.
More info...
Data imputation using machine learning in order to improve echocardiographic assessment of
cardiac mechanics – #71
Promotors: Frederik Maes , Jan D'hooge – Mentor: Mahdi Tabassian
Background
Echocardiographic deformation imaging is the modality of choice for the noninvasive assessment of regional heart function. Quantification of regional
contractile function of the left ventricle (LV) is performed by dividing it into segments and measuring the deformation of each segment (Fig1.a). Like any
imaging modality, echocardiographic deformation imaging has its own limitations and technical challenges. One of these challenges is that the acquired
images should be of good quality for the deformation curves to be reliable. Poor image quality, due to artifacts, noise, aliasing etc., may lead to
deformation curves that are artifactual and not physiologically meaningful (Fig1.b). Such traces are not of use and cannot be analyzed by the
cardiologist or an automatic classification algorithm. Even worse, they can appear pathophysiologic and thereby cause confusing and misdiagnosis.
Moreover, acquiring images of all heart segments is not always possible. This can happen because of shadowing by ribs or lung tissue (Fig1.c). As a
consequence, deformation curves may be missing for some of the LV segments.
Content
The objective of the current project is to deal with the problem of missing or unreliable deformation curves of the heart via ‘machine learning’. More
specifically, the suitability of different ‘data imputation’ techniques [1] will be examined to: (i) estimate the artifactual/missing deformation curves (e.g.
using the Knearest neighbor imputation technique that has been successfully implemented by our group [2] (Fig1.d)) and, (ii) measure reliability of the
available deformation curves. Different ‘clinical markers’ that are usually used by the cardiologist to quantify heart function will be extracted from the
imputed curves obtained in part (i) and subsequently compared with those taken from the original curves. Also, a ‘reliability map’ will be generated for
the deformation curves to demonstrate the likelihood of being affected by artifacts, noise, etc., using the outcome of the analysis performed in part (ii).
In the next stage, an automatic classification system will be trained and tested with and without the use of the reliability map to verify whether
identification of potential measurement errors in the deformation data has a significant impact on the classification results or not.
Practical
The thesis project is conducted in close collaboration with the group of Cardiovascular Imaging and Dynamics of KU Leuven. The aforementioned tasks
will be performed using four invivo databases available in the lab. Two of these databases include segmental deformation curves acquired from 110
patients with myocardial infarction and the other two contain data of 100 normal control subjects. Data analysis will be performed using Matlab and the
student can use the scripts that have been developed inhouse for data preprocessing, imputation and classification.
Prerequisite
Experience in machine learning and statistical pattern recognition as well as having solid programming skills will help the candidate to accomplish the
current project.

References
[1] Troyanskaya O, Cantor M, Sherlock G, Brown P, Hastie T, Tibshirani R, Botstein D, Altman RB, “Missing value estimation methods for DNA
microarrays”, Bioinformatics, 17(6):5205, 2001.
[2] Tabassian M, Alessandrini M, Jasaityte R, De Marchi L, Masetti G, D'hooge J, “Handling missing strain (rate) curves using Knearest neighbor
imputation”, IEEE Int Ultrason Symp (IUS), 14, 2016.
Location
4. Biomedical Instrumentation
Bladder volume estimations using inertial sensors – #42
by Weydts Tristan – last modified Mar 30, 2017 15:10
Discipline: Biomedical Instrumentation
Promotor: Bob Puers – Mentor: Tristan Weydts
An implantable sensor network of accelerometers was designed to learn more about bladder wall movements. The system has multiple stretchable arms
and can enclose the bladder during its stretching and shrinking. The existing system can be adapted to estimate the volume of the bladder using inertial
sensors such as accelerometers, gyroscopes, magnetometers, … The student will search for optimal conditions to estimate the bladder volume. The
amount of measurement points/sensors, the types of sensors and the readout algorithm will have to be considered. In this thesis a mathematical model
will be developed, followed by an implementation using commercial hardware and accompanying software. Real data is available for verification of the
model. Afterwards there is the possibility to actually implant the system and test it.
Additional Info:
Possible attendance of animal experiments (operating room).
Microvalve manifold for High Pressure Liquid Chromatography (HPLC) application. – #46
by Ceyssens Frederik – last modified Mar 29, 2017 19:22
Promotor: Bob Puers – Mentors: Frederik Ceyssens , deirdre cabooter (dept of pharmacy)
This project is part of our research on improving the sensitivity and throughput of chemical analyses based on the HPLC technique by miniaturization
and parallelization.
HPLC is extremely useful and widely used in for example drug development, medical tests, pharmaceutical analysis, organic chemistry, etc.
In HPLC, switching valves are frequently used to increase the number of stationary phase types that can be evaluated without manual interference.
Commercially available switching valves currently allow for a restricted number and combination of stationary phases only. The goal of this project,
therefore is to investigate whether the flexibility of these switching valves can be increased by developing new valve concepts based on
microtechnology.
Challenges are the fact that these valves should be able to withstand high pressures (up to 1000 bar), should be compatible with typical solvents used
in liquid chromatography (acetonitrile, methanol) and the sealing should be reversible. A possible approach is to base the valves on phase change
actuators.
It is a nice engineering topic that involves doing microfluidic design, backofthe envelope calculations, material selection, fabrication of simple test
structures using rapid prototyping and other techniques, and testing in our lab. Then, a more advanced prototype can be built in the Corelab cleanroom
using lithographic fabrication technologies. As this is a collaboration with the department of pharmaceutical sciences, it also involves a lot of user
interaction and testing of the prototypes at the pharmaceutical department.
For more info, please contact the mentortobe at 0494 465566 or by email.
Development of a differential, highly sensitive pressure sensor for the cochlea – #48
by Puers Bob – last modified Mar 30, 2017 01:30
Promotors: Bob Puers , Nicolas Verhaert (UZ Leuven) – Mentor: Grim Keulemans
Sound leads to pressure differences inside of the human cochlear fluids, and thus to hearing. Measuring these pressure differences allows for
disentangling the most optimal inner ear stimulation and the development of new hearing acoustical implants.

In this thesis, an extremely miniaturized pressure sensor must be developed, that is capable of being connected to the two cochlea channels. A
selection of the most suitable sensor principle is the first task. The sensors must be robust, easily calibrated and connected to a validated and
transportable signal conditioner. The sensor’s frequency spectrum must correlate to the human hearing, i.e. from 0.1 to 10 kHz, and be capable to
measure sound pressure levels from 40 to 100120 dB SPL, corresponding to 0.002 Pa to 2Pa (!!). Preferably, the sensors can be used for multiple
measurements without the need for recalibration.
Once validated, the sensors will be inserted through small openings inside of the cochlea with a max diameter of 300 µm and an insertion depth of 100
to 150µm, as shown in the figure 1.
Resorbable materials for temporary reinforcement of high resolution neural implants – #47
by Ceyssens Frederik – last modified Mar 29, 2017 19:34
Promotor: Bob Puers – Mentor: Frederik Ceyssens

Lately, evidence has been found that it is possible to implant devices in the brain without any scar tissue formation over the long term, as long as the
materials are inert and the device is extremely flexible and can move with the brain.

Thus, the dream of longterm, high resolution interfacing directly with the brain could be reality. Still, a major practical hurdle is how to insert very
compliant structures in the brain.

An option would be to make such soft, compliant neural implants and temporarily reinforce them with a material that dissolves over a period of days or a
few weeks at maximum.

In this thesis, you would select a few resorbable materials based on expected biocompatibility, resorption time and the suitability for micromachining and
integration in thin film neural implants.

Then, you would built test structures that include these materials, using micromachining techniques. Finally, an in vivo test is foreseen to gather hard
evidence of the material's performance.
Design of a selfattaching gelfree microneedle electrode for EEG recordings – #49
by Bertrand Alexander – last modified Mar 30, 2017 15:10
Promotors: Bob Puers , Alexander Bertrand – Mentors: Frederik Ceyssens , Abhijith Mundanad (KU Leuven)
Electroencephalography (EEG) measures electrical brain activity through electrodes that are attached to the scalp, and is currently viewed as the most
promising noninvasive modality towards chronic 24/7 neuromonitoring in daily life. However, current EEG systems use bulky headsets in order to press
the electrodes against the scalp, which is uncomfortable and stigmatizing (see Fig. 1). Furthermore, EEG typically requires skin abrasion to remove the
upper skin layer, and a lowimpedance connection is obtained by applying electrolytic gel between the electrodes and the skin. This requires a long set
up time by an experienced person, and also hampers a longterm recording as the gel dries out over time.
The goal of this thesis is to design a new type of microneedle EEG electrode with 2 gamechanging features:
1) It penetrates the resistive upper skin layer using microneedles (see Fig. 2) to reach the conductive epidermal layer, eliminating the need for gel and
skin abrasion.
2) It locks itself onto the skin (even at hairy regions). This avoids the need for bulky headsets to put mechanical pressure on the electrodes to keep
them in contact with the skin.
These two features would be a major step forward towards chronic EEG electrodes. Also for shortterm recordings, they will allow to substantially
decrease the setup time and improve comfort.
Project
After a brief literature study, we will make a new microneedle electrode design and fabricate them in an ISO class 6 clean room, inspired by the design
in [1]. We will test their ability to lock onto the (hairy) scalp and alter the design if necessary. Validation will be based on standard braincomputer
interfacing (BCI) experiments with inhouse software.
We expect a strong motivation and commitment. If you plan to select this thesis, please contact us beforehand.
Promoters:
Prof. Robert Puers (KU Leuven, ESATMICAS) ([email protected])
Prof. Alexander Bertrand (KU Leuven, ESATSTADIUS) ([email protected])
Supervision:
Electrode design: Dr. Frederic Ceyssens (KU Leuven, ESATMICAS) ([email protected])
BCI experiments: Abhijith Mundanad (KU Leuven, ESATSTADIUS) ([email protected])
Work load: literature 10% implementation 35% validation 55%
5. Biomaterials
Exploring the potential of shape memory polyurethane foam for biomedical applications – #65
Discipline: Biomaterials
Promotor: Luc Labey – Mentors: Luc Labey , Carl de Smet (Noumenon)
Background:
Shape memory polyurethane (SMPU) is a material which changes shape upon heating (much like shape memory alloys, though the physical mechanism
is different). Because of their attractive combination of properties (lightweight yet relatively stiff and strong), foams made from this material have already
been used (or are under investigation) for applications in furniture (Figure 1) and the car industry. Obviously, these properties may also lead to interesting
biomedical applications.
A potential application in the biomedical field that immediately springs to mind is for fixation purposes. SMPU foams might offer interesting alternatives
for casts or splints for instance. The advantage of SMPU compared to the conventional casts is that they can be taken off and applied again more
easily, which is advantageous from a hygienic perspective. Contrary to splints, the SMPU foams will fit the shape of the limb better and offer more
support.
During some scans (such as MRI) or radiotherapy, the patient’s body needs to be immobilised for a certain amount of time. Also in this situation, SMPU
may be an alternative for the currently used techniques (for instance the use vacuum immobilisation mats during MRI scanning).
Finally, external devices to support or assist the body (such as in orthoses or prosthesis sockets) require a good fit to the limb to function properly.
Since they can be made to the shape of the body, SMPU’s may be interesting for this kind of applications too.
The objective of this thesis is to explore the above mentioned examples (and others, if they come up) more thoroughly and develop demo material for
those that are interesting enough to be taken into further consideration. These demos will be used to illustrate the potential of SMPU during discussions
with partners (both clinical and industrial) for further development.
An obvious first step will be a (literature) study to identify shortcomings in existing techniques that may be alleviated by the use of SMPU. Secondly, for
those applications for which the use of SMPU reveals true benefits, an appropriate demo will be developed to illustrate exactly this benefit. This will
require the establishment of required material properties (density, stiffness, strength, transformation temperature,…) and shape transformation for the
device (or assembly). In a final step, the demo(s) will be validated to make sure that they perform as intended.
More info...
6. Tissue Engineering
No proposals available for this discipline
Copyright © KU Leuven Werktuigkunde | reacties op de inhoud: POC BMT
Realisatie: ICT Werktuigkunde | Versie: 0.3.0 | Disclaimer

BME Masters Theses 2017-2018

Uploaded by

Copyright:

Available Formats

BME Masters Theses 2017-2018

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

BME Masters Theses 2017-2018

Uploaded by

Copyright:

Available Formats

What are some potential applications of shape memory polyurethane foams discussed in the document?

What are the objectives of exploring shape memory polyurethane foams for biomedical applications?

3/30/2017 BME

Promotors: Jos Vander Sloten , Nele Famaey – Mentors: Markos Kapeliotis , Heleen Fehervary

Promotors: Jos Vander Sloten , Nele Famaey – Mentor: Patricia Silvana Da Torre Lopes (KU Leuven Materialise)

Promotors: Jos Vander Sloten , Guido de Bruyne (Lazer Sport) – Mentors: Dimitris Zouzias (Ku Leuven / Lazer Sport) , Michel Woering ,

Promotors: Hans Van Oosterwyck , bart Smeets (MeBios) – Mentors: Tommy Heck , Bart Smeets (MeBios)

Promotors: Sabine Van Huffel , Alexander Caicedo Dorado , Borbála Hunyadi – Mentor: Simon Van Eyndhoven

Promotor: Alexander Bertrand – Mentor: Abhijith Mundanad (ESAT)

Promotors: Sabine Van Huffel , Gunnar Naulaers , Alexander Caicedo Dorado – Mentors: Mario Lavanga , Ofelie De Wel

We recently developed a proofofconcept that demonstrates that deep learning

You might also like

BME Masters Theses 2017-2018

Uploaded by

Copyright:

Available Formats

BME Masters Theses 2017-2018

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

BME Masters Theses 2017-2018

Uploaded by

Copyright:

Available Formats

What are some potential applications of shape memory polyurethane foams discussed in the document?

What are some potential applications of shape memory polyurethane foams discussed in the document?

What are the objectives of exploring shape memory polyurethane foams for biomedical applications?

What are the objectives of exploring shape memory polyurethane foams for biomedical applications?

3/30/2017 BME

Promotors: Jos Vander Sloten , Nele Famaey – Mentors: Markos Kapeliotis , Heleen Fehervary

Promotors: Jos Vander Sloten , Nele Famaey – Mentor: Patricia Silvana Da Torre Lopes (KU Leuven ­ Materialise)

Promotors: Jos Vander Sloten , Guido de Bruyne (Lazer Sport) – Mentors: Dimitris Zouzias (Ku Leuven / Lazer Sport) , Michel Woering ,

Promotors: Hans Van Oosterwyck , bart Smeets (MeBios) – Mentors: Tommy Heck , Bart Smeets (MeBios)

Promotors: Sabine Van Huffel , Alexander Caicedo Dorado , Borbála Hunyadi – Mentor: Simon Van Eyndhoven

Promotor: Alexander Bertrand – Mentor: Abhijith Mundanad (ESAT)

Promotors: Sabine Van Huffel , Gunnar Naulaers , Alexander Caicedo Dorado – Mentors: Mario Lavanga , Ofelie De Wel

We recently developed a proof­of­concept that demonstrates that deep learning

You might also like

Promotors: Jos Vander Sloten , Nele Famaey – Mentor: Patricia Silvana Da Torre Lopes (KU Leuven Materialise)

We recently developed a proofofconcept that demonstrates that deep learning