Computer Methods and Programs in Biomedicine 208 (2021) 106200

Contents lists available at ScienceDirect

Computer Methods and Programs in Biomedicine

journal homepage: www.elsevier.com/locate/cmpb

Delivering computationally-intensive digital patient applications to the

clinic: An exemplar solution to predict femoral bone strength from CT
data
I. Benemerito a,b,∗, W. Griffiths a,b, J. Allsopp c, W. Furnass d, P. Bhattacharya a,b, X. Li a,b,
A. Marzo a,b, S. Wood c, M. Viceconti f,g, A. Narracott a,e
a
INSIGNEO Institute for in silico Medicine, The University of Sheffield, UK
b
Department of Mechanical Engineering, The University of Sheffield, UK
c
Sheffield Teaching Hospital Foundation Trust, Sheffield, UK
d
Department of Computer Science, The University of Sheffield, UK
e
Department of Infection, Immunity & Cardiovascular Disease, The University of Sheffield, UK
f
Department of Industrial Engineering, Alma Mater Studiorium, University of Bologna, Italy
g
Medical Technology Lab, IRCSS Istituto Ortopedico Rizzoli, Bologna, Italy

a r t i c l e i n f o a b s t r a c t

Article history: Background and objective: Whilst fragility hip fractures commonly affect elderly people, often causing per-
Received 23 December 2020 manent disability or death, they are rarely addressed in advance through preventive techniques. Quan-
Accepted 19 May 2021
tification of bone strength can help to identify subjects at risk, thus reducing the incidence of fractures
in the population. In recent years, researchers have shown that finite element models (FEMs) of the hip
Keywords: joint, derived from computed tomography (CT) images, can predict bone strength more accurately than
Bone strength estimation other techniques currently used in the clinic. The specialised hardware and trained personnel required
Clinical workflow to perform such analyses, however, limits the widespread adoption of FEMs in clinical contexts. In this
Decision support manuscript we present CT2S (Computed Tomography To Strength), a system developed in collaboration
XNAT
between The University of Sheffield and Sheffield Teaching Hospitals, designed to streamline access to
Finite element modelling
this complex workflow for clinical end-users. Methods: The system relies on XNAT and makes use of cus-
Open source software
tom apps based on open source software. Available through a website, it allows doctors in the healthcare
environment to benefit from FE based bone strength estimation without being exposed to the technical
aspects, which are concealed behind a user-friendly interface. Clinicians request the analysis of CT scans
of a patient through the website. Using XNAT functionality, the anonymised images are automatically
transferred to the University research facility, where an operator processes them and estimates the bone
strength through FEM using a combination of open source and commercial software. Following the analy-
sis, the doctor is provided with the results in a structured report. Results: The platform, currently available
for research purposes, has been deployed and fully tested in Sheffield, UK. The entire analysis requires
processing times ranging from 3.5 to 8 h, depending on the available computational power. Conclusions:
The short processing time makes the system compatible with current clinical workflows. The use of open
source software and the accurate description of the workflow given here facilitates the deployment in
other centres.
© 2021 The Author(s). Published by Elsevier B.V.
This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/)

1. Introduction elling to support academic research in several clinical domains in

recent years [1]. An important class of these applications are Dig-
The constant improvement of computational capabilities has led ital Patient or Digital Twin solutions, where the data of an in-
to a surge of clinical and biological applications of computer mod- dividual patient are used to inform a computer model that pro-
vides the clinical specialist with information useful to personalise
∗
the management of that patient. These computational models al-
Corresponding author at: The University of Sheffield, Pam Liversidge Building,
Room E09, Mappin Street, Sheffield, S1 3JD, UK.
low the study of scenarios otherwise difficult to investigate exper-
E-mail address: i.benemerito@sheffield.ac.uk (I. Benemerito). imentally. Their deployment, however, requires specialised hard-

https://doi.org/10.1016/j.cmpb.2021.106200
0169-2607/© 2021 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/)
I. Benemerito, W. Griffiths, J. Allsopp et al. Computer Methods and Programs in Biomedicine 208 (2021) 106200

ware and software, trained personnel and high performance com- the individual components and how they are integrated to provide
puting (HPC) systems, which are expensive to install and maintain a complete service.
and whose use is challenging for people without specific training.
These factors limit the widespread adoption of these technologies
2.1. Description of the workflow
in clinical practice.
A solution to this problem is to hide the technical part within
Figure 1 shows the architecture and schematic of the operation
a wrapper, thus allowing clinicians to access the computational
of the workflow, which operates over the Hospital (STH) and Uni-
models without needing to build them. This approach delivers HPC
versity (TUOS) networks.
resources to the clinic and improves the clinical translation of
The workflow operates upon the cooperation of a number of
biomedical research. A number of platforms have been developed
University developed applications (CT2SWebApp, django-multisim,
over the years for sharing medical data across institutions (Som-
DicomAnonymiser, DataExchange, AMPQClient), either based on
nonetz for sharing polysomnographic data [2], Origa-light for im-
existing open source software or written from scratch. Figure 2
ages of the retina [3]) and for facilitating the retrieval [4] and re-
shows a simplified representation of the workflow, while Fig. 3
construction [5] of medical images, but specific platforms for clini-
shows the individual components of the service and their mutual
cal decision support have not seen widespread uptake. Notable ex-
relationship.
amples include systems for detection of heart conditions (VIRTU-
A patient reaches the hospital and undergoes a bilateral full
heart [6] and HeartFlow [7]), or the use of finite element (FE) mod-
femoral CT scan at 100 kV. The amperage, between 80 and 200
elling for estimating the risk of hip fracture, an application that has
mA, is chosen according to the patient’s weight. This corresponds
recently been approved by the U.S. Food and Drug Administration
to an effective radiation dose ranging from 1.5 to 6.2 mSv [19].
as a class II medical device (VirtuOst [8]). In all these cases, clini-
The CT machine has been calibrated during the deployment of the
cal staff are required to electronically transfer a set of medical im-
workflow using an European Spine Phantom [20]. However, as the
ages to operators that use these images as input for computational
calibration is asynchronous, the use of phantom is not required
models.
during the scanning procedure. Following scan acquisition, the im-
Fragility hip fractures commonly affect elderly people, espe-
ages are automatically saved into the “Picture Archive and Commu-
cially post-menopausal women [9], and constitute a significant
nication System” (PACS) [21] according to standard clinical practice
burden on the healthcare system, with an estimated cost of £1.1
in UK hospitals. In accordance with the guidelines set by the In-
billion in the UK alone [10]. They are usually caused by osteo-
formation Governance of the hospital, any information related to
porosis, a disease which weakens the bones, including the femur,
the patient’s identity is removed from the headers of the DICOM
and makes them more prone to fracture [11]. Patients at risk can
images. The system also generates a unique identifier that is asso-
be identified using dual-energy X-ray absorptiometry imaging to
ciated with the simulation request. The information that links the
measure the areal bone mineral density (BMD). This can then be
identifier to the patient’s identity, and the specific request for sim-
related to the bone strength, defined as the force needed to frac-
ulation service remains stored safely behind the hospital firewall.
ture a bone [12], which is considered a biomarker for the risk of
To request a CT2S analysis, firstly a member of clinical staff logs
fracture [13,14]. A recent study has shown that this can achieve a
into the Hospital PACS Image Exchange system (PIX) and completes
predictive accuracy of 75% [15].
an image transfer request, specifying the fields “Subject ID” and
Alternatively, researchers have developed methods to use med-
“Subject Label”. The PIX system transfers the DICOM files from the
ical images, typically in DICOM format [16] from quantitative com-
Hospital PACS system to the University XNAT. No personal data or
puted tomography (CT), as a starting point for building person-
identifier are shared across institutions. Secondly, the clinician logs
alised 3D FE models of the femur [17] and estimating its strength
onto the CT2S website and completes a submission form, ensuring
[15,18]. Over the years our group has developed a workflow to pre-
that the field “Job Ref.” matches either the “Subject ID” or “Subject
dict bone strength using clinically acquired CT images, and vali-
Label” submitted to PIX.
dated it on a cohort of postmenopausal women. This approach can
CT2SWebApp, which manages the CT2S website by extending
reach accuracy above 80% [18], but is not yet commonly used in
and overriding the django-multisim library, receives the metadata
the hospitals because it requires a sophisticated set of software,
of the submitted job, validates the submission and updates the job
trained operators and HPC resources typically not available in clin-
status as the workflow progresses. It is also responsible for locating
ical settings.
the images in the database, and coordinates the communication
To address this issue, INSIGNEO Institute for in silico medicine
of metadata across the apps django-multisim and DataExchange,
at The University of Sheffield (TUOS) and Sheffield Teaching Hos-
as well as the image transfer to ShARC, the Tier 3 HPC system
pital (STH) have developed a web service, named “CT to Strength”
at TUOS [22], through the app DicomAnonymiser. The images are
or CT2S for short, and made available at https://ct2s.insigneo.org/
made available to the operator, who can now perform the FE analy-
ct2s/. Using the system, doctors in hospitals can request the anal-
sis on the HPC cluster. Results from the analysis are then uploaded
ysis of a set of CT scans and obtain a comprehensive report con-
on the website and made available to the referring clinician in a
taining the key predictive information. Overall, the service aims to
PDF document.
improve the clinical uptake of FE-based strength estimation with-
out exposing the complexity of the underlying HPC environment,
which is shielded from the clinician by a user friendly interface. 2.2. Infrastructure

2. Materials and methods 2.2.1. XNAT

The image transfer elements of the workflow are provided by
The workflow behind the CT2S service is built over two sep- XNAT 1.7 [23], an open source research platform for image-based
arate networks, the first one located within the Sheffield Teaching biomedical research. XNAT allows users to collaborate on projects
Hospital (STH), referred to as the Hospital network, and the second and share data and metadata using a database environment. For
one within The University of Sheffield (TUOS). The entire workflow DICOM images most of the metadata is directly inherited from the
can be split into two parts: infrastructural and modelling. image itself through their headers, which contain details on the
In the next sections we will first present a high level descrip- image such as resolution, slice thickness, location and orientation
tion of the workflow, followed by a more in depth description of of reference frames, and information on the patient such as name,

2
I. Benemerito, W. Griffiths, J. Allsopp et al. Computer Methods and Programs in Biomedicine 208 (2021) 106200

Fig. 1. Operation of CT2S workflow from initial clinician request to final reporting of bone strength assessment. The red arrows represent the DICOM image being pushed
from the Hospital (STH) PACS system to the University (TUOS) XNAT through PIX. The images are then transferred to the University HPC system (ShARC) and downloaded by
the operator. Yellow arrows indicate the flow of non-image data. (For interpretation of the references to color in this figure legend, the reader is referred to the web version
of this article.)

and “Subject Label”. The generic identifiers are shared with TUOS,
while the mapping between them and the metadata is kept within
the STH network. This process ensures that no information allow-
ing identification of the patient leaves the Hospital network at any
time.

2.2.2. PIX
The PIX system provides an interface between the Hospital
hosted PACS and the University hosted XNAT. Before initiating a job
submission, the clinician searches PACS using the patients hospital
number to find the imaging studies for the analysis. Before initiat-
ing the transfer, the system requires to enter the study ID for the
patient on the trial and to specify the destination project, as the
XNAT server contains many projects for a wide range of research
applications (Fig. 4).
Technically the following steps take place:

Fig. 2. Simplified representation of the workflow. Each box is coloured according • The transfer request gets added to the job queue and waits to
to the app that performs the action. (For interpretation of the references to color in be started.
this figure legend, the reader is referred to the web version of this article.) • The job starts and a “DICOM Send” request is sent to the PACS
system for the desired studies to be transferred to the PIX
server. PIX uses the Orthanc project as its DICOM node.
sex and identification number, specifically “Subject ID” and “Sub- • The images are then anonymised with the requested trial ID
ject Label”. During the de-identification step, the clinical personnel and all potentially identifiable data within the DICOM image
generates generic identifiers to replace the metadata “Subject ID” headers is removed.

Fig. 3. Apps involved in the workflow and their relationship. Green boxes are the apps at the core of the CT2S workflow. CT2SWebApp receives the job metadata from the
website and directs the operation of the apps responsible for job validation and image transfer. The red arrow represents the DICOM image being copied from the XNAT
database to the University HPC system (ShARC), where it is made available to the operator. (For interpretation of the references to color in this figure legend, the reader is
referred to the web version of this article.)

3
I. Benemerito, W. Griffiths, J. Allsopp et al. Computer Methods and Programs in Biomedicine 208 (2021) 106200

Fig. 4. Overview of the image transfer process between Hospital PACS and University (TUOS) XNAT, managed by PIX. DICOM images can be uploaded to PACS from PC, DVD
or directly from the imaging machine. PIX initiates a transfer request from PACS to its server and, following the successful anonymisation of the DICOM images, sends the
images to the XNAT server. Red arrows indicate the flow of image data, the yellow arrow is the flow of the configuration file for batch processing, blue arrows represent
communication internal to the PIX system. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

• The anonymised images are uploaded to the requested XNAT DicomAnonymiser (DA) is a TUOS developed utility, written
server via REST transactions [24]. in C# [28]. Despite its name, the app does not perform de-
• The images are deleted from the PIX server and the job is identification, as this is done within the Hospital network. This
flagged as complete. means that the University does not manage any personal data of
the patients and only handles de-identified images. DA is used to
The PIX service also has a batch mode where CSV configuration
receive instructions to transfer DICOM images downloaded by DE
files containing the information can be uploaded and thousands of
to the shared drive hosted on ShARC and notifies CT2WebApp of
imaging studies can be provisioned automatically.
the successful transfer of the images. Communication between DA
and other apps is facilitated using the messaging client AMQP-
2.2.3. Components of the workflow: django-multisim, DataExchange,
Client.
DicomAnonymiser, AMPQClient
AMPQClient is the RabbitMQ [29] client which allows the apps
The django-multisim (DM) is a TUOS developed app, based on
involved in the workflow to send and receive messages. RabbitMQ
Django [25] and integrated with other services provided by IN-
is open source message-broker software that allows communica-
SIGNEO. Django is a free, open source web framework written
tion between apps using different communication protocols. It is
in Python [26], which facilitates the creation of database-driven
written in Erlang [30], but allows for integration with clients writ-
websites providing a variety of services and interface capabilities.
ten in other languages. Exploiting this, the AMPQClient is written
DM and CT2SWebApp, which also provide a constantly running
in Python and added as a library to the other apps of the workflow.
daemon (known as “ct2s-main”) on the CT2S website virtual ma-
chine in order to communicate with RabbitMQ, intervene in all the
phases of the workflow. DM validates a submitted job by checking 2.2.4. Components of the workflow: CT2SWebApp
that all the fields in the submission form have been correctly com- CT2SWebApp, written in Python 3 and based on Django, is cen-
pleted using a JSON schema. These fields are job ref. (must match tral to the entire workflow, as it provides the CT2S web site and al-
the patient ID supplied to PIX), age, gender, weight, height, project, lows users to interface with the front end of the service to initiate
ethnicity. After the validation phase, DM retrieves the study meta- requests to perform a CT2S analysis. CT2SWebApp interfaces with
data using the XNAT API and assigns the job to an operator, who is the other apps and orchestrates their operations using the mes-
notified by email. Finally, DM uses JavaScript [27] to generate the saging functionalities provided by AMPQClient. A user account is
final PDF report. The clinician is notified by email after the creation required for submitting jobs to the platform. To create and man-
of the PDF file. age users, CT2SWebApp extends the DM library: users are classi-
On the University side, DICOM images are initially stored in the fied into three categories (administrator, operator and user), each
University XNAT database. They are made available to the opera- with different privileges. All the classes can submit jobs, whose va-
tor through a series of steps, the first of which is performed by lidity is assessed by DM by ensuring that all information necessary
the app DataExchange (DE), developed at TUOS in Python 3.6. Fol- for the analysis has been provided. Following a job submission,
lowing job validation, DM sends to DE the metadata of the DICOM CT2SWebApp, with DM functionality, validates the submission and,
image needed for the study. After DE locates the study within the following validation, monitors the University XNAT database to lo-
University XNAT database, it downloads it to its own virtual ma- cate the images associated with the job. This is done by check-
chine hosted within the University network. Following the down- ing whether the field “Job Ref.” in the submission script matches
load, DE sends a message to DicomAnonymiser via the AMQP- either the “Subject ID” or “Subject Label” in the XNAT metadata.
Client. Furthermore, CT2SWebApp communicates via email with the end

4
I. Benemerito, W. Griffiths, J. Allsopp et al. Computer Methods and Programs in Biomedicine 208 (2021) 106200

user, who is notified every time the job changes status. Possible As shown in Fig. 5, the end user is requested to enter the age, gen-
job statuses are: der, weight and height of the subject. The ethnicity can also be en-
tered, although is an optional field. The fields “Project” and “Eth-
• submitted: waiting for the corresponding study to be detected
nicity” are a reference for the end user. The field “Job Ref.”, instead,
on the University XNAT service
is fundamental for the correct execution of the workflow and must
• pending: waiting for job data to be sent to DE after XNAT study
match either the “Subject ID” or “Subject Label” fields in the XNAT
metadata is retrieved
database.
• processing: job data has been sent to DE
• failed: the CT2S analysis was unsuccessful
3.2. PDF report
• open: anonymised scans are deposited in the shared drive
• done: results of the CT2S analysis are added to the job form
After downloading the images, building the model and perform-
• discarded: the job has not failed but CT2S analysis will not be
ing the analysis, the operator generates the final report containing
completed for this job. This can happen because of multiple
the information relevant for clinical assessment of bone strength.
reasons (e.g. the input data is incorrect)
Figure 6 shows an example report.
After the analysis is complete, the operator submits to the web- The report contains information about the job, such as its “Job
site the data for the creation of the PDF report. CT2SWebApp ex- Ref.”, the date of submission and the job number, as well as the
tends DM to create the PDF report and notifies the end user of the name of the operator who performed the analysis. Minimum and
termination of the job. maximum values of the strength of the femur are reported, to-
gether with a graphical representation of the strain field, for both
2.2.5. Storage stance and fall scenarios.
The operator does not have direct access to the images, and
is not able to download them from the XNAT database. Instead, 3.3. Performance
images are uploaded to a job-specific folder with restricted access
rights on ShARC. Table 1 presents the timing of the workflow. The image trans-
fer between the Hospital PACS and University XNAT is managed
2.3. Modelling automatically within the Hospital network initially by PIX, and
then by internal XNAT functionality. Following job submission,
2.3.1. ITK-Snap, Bonemat, ICEM, ANSYS and MATLAB the CT2SWebApp continuously scans the University XNAT database
The modelling elements of the workflow focus on providing an to locate the scans associated with the requested job. Likewise,
estimation of the strength of an individual’s femur under a se- ct2s-main (provided by django-multisim and CT2SWebApp) and
ries of loading conditions, using a subject-specific FE model. This DataExchange run as daemons to minimise the waiting time as-
step involves the use of ITK-Snap [31] for image segmentation to sociated with the transfer from University XNAT to HPC system
isolate the bone surface, ICEM CFD 15 (ANSYS Inc, PA, USA) for (ShARC). Wall clock time for FE analysis is reported for both the
meshing the segmented bone geometry using tetrahedral elements, parallel and serial runs. Despite most of the time being spent on
Bonemat [32] for personalising the bone material properties, and the semi-automatic part of the workflow, more specifically during
ANSYS 17 (ANSYS Inc, PA, USA), running on ShARC, for perform- the segmentation of bone geometry, the total time for completing
ing the simulations. The material is assumed to be linearly elastic the workflow is less than a day, so clinical staff can review the data
and heterogeneous, with each element being assigned its Young’s before the next appointment with the patient.
modulus according to the corresponding attenuation value in the
CT image [18]. A total of 28 load simulations are performed to 4. Discussion
comprehensively investigate the bone response under several fall
configurations.The FE model used in CT2S implements the multi- The CT2S workflow has been developed with the clear target
point constraints method to describe the kinematics of the fall, and of integration within clinical practice, and aiming to provide clini-
the contact between the femur and the ground surface is mod- cians with state-of-the-art predictive tools. The workflow has been
elled as large sliding frictionless contact. The origin of the coordi- developed in alignment with General Data Protection Regulations
nate frame is located at the centre of the femoral head, with the principles, which require that subjects must consent to the col-
axes defined using anatomical landmarks. The maximum principal lection and processing of their personal data [33]. When patients
strain criterion has been used to estimate the failure load [17]. are scanned for CT2S analyses, their personal data do not leave the
Detailed methodologies have been published in previous papers Hospital network and the only metadata shared across institutions,
[15,18]. Each load simulation requires 96 GB of memory shared “Subject ID” or “Subject Label”, do not allow the identification of
across three processors in an OpenMP session. USFD has the capa- the patient. Patients also sign a consent form for their anonymised
bility of running the entire batch of 28 simulations simultaneously, CT scans to be processed at TUOS.
but the analysis can also be run serially. Results are post processed In addition to prediction capabilities, economic considerations
in MATLAB 2018a (MathWorks, Natick, MA) with custom written are also critical for an effective translation of modelling approaches
scripts. to clinical practice and for a multicentre deployment of the work-
flow. Costs associated with the CT2S workflow include the scan-
3. Results ning of the patient, the training of the operator, the purchase and
maintenance of the required HPC resources and software licenses.
3.1. Upload interface The current version of CT2S adopts commercial software for mesh-
ing and FE simulations. To reduce costs and encourage a wider dif-
The end user interacts with the system through PIX, which is fusion of the workflow it would be possible to replace them with
a system internal to the Hospital, and the website https://ct2s. open source meshing and FEA packages, although this could re-
insigneo.org/ct2s/, which is managed by CT2SWebApp. Following quire additional effort in terms of code verification and validation
log-in to the CT2S website, the end user has the possibility to sub- to ensure the accuracy of the whole workflow. The main techni-
mit a new job or to see the details of the jobs already submitted. cal limitation of the workflow is the semi-automatic segmentation
New jobs are submitted through a dedicated submission platform. pipeline, which relies heavily on human intervention while the rest

5
I. Benemerito, W. Griffiths, J. Allsopp et al. Computer Methods and Programs in Biomedicine 208 (2021) 106200

Fig. 5. Submission form from the website. After the DICOM images have been uploaded to PIX,the end user fills the submission form to request the job.

Table 1
Timescales for operations of the CT2S workflow.

Workflow action Duration Comments

Upload images to PIX 10 min Depends on time of day and load on

the clinical PACS system
Transfer from PIX to University 5 min Depends on number of slices in the
XNAT CT image
Login to CT2S web portal Less than 1 min
Submit job 5 min
Validate job 1 min
CT2SWebApp locates the scans 2 min Assumes scans are already present
on TUOS XNAT (waits for up to 2
days if not)
Retrieve XNAT metadata 5 min
DataExchange downloads scans to 3 min Depends on download size
virtual machine
DicomAnonymiser copies scans to 3 min Depends on scan file sizes
shared drive
Operator downloads scans 5 min Depends on number of slices in the
CT image
Image segmentation 2h Depends on number of slices in the
CT image
3D meshing 5 min
Material property estimation 3 min
Solve FEM in ANSYS From 10 min to 4.5 h 10 min is the time for simulating a
single loading direction. It the 28
loading configurations are simulated
in parallel then the total computing
time is 10 min. Sequential run
requires 4.5 h
Download results Less than 10 min
Post processing 10 min
Produce final report 5 min
TOTAL TIME From 3.5 to 8 h 3.5 h is the total workflow time for
parallelised FE analysis. If loading
simulations cannot be performed in
parallel that the total time is
approximately 8 h

of the workflow is automated thanks to existing software pack- process [34]. These techniques work by warping template meshes,
ages and scripts that link together the various software compo- taken from anatomical atlases, to the patient’s femur in order to
nents, and weighs considerably on the determination of the cost fit relevant anatomical features. Alternatively, recent developments
of the service. The image segmentation is performed using open in machine learning have led to tools that can perform accurate
source software, which has been a deliberate design choice to re- segmentation in minutes, and in some cases these tools have been
duce costs and foster the wider deployment of the platform. Elastic embedded within commercial software such as Simpleware (Syn-
registration algorithms can assist and speed up the segmentation opsis Inc, Mountain View, CAL). Although the adoption of commer-

6
I. Benemerito, W. Griffiths, J. Allsopp et al. Computer Methods and Programs in Biomedicine 208 (2021) 106200

Fig. 6. Snapshot of the PDF report made available to the end user. In the top half, details on the job are given. In the lower half, the predicted strength is reported for stance
and falling and, for each configuration, the corresponding deformation field is shown.

cial software for automatic segmentation would certainly reduce same statistical power, which implies a significant reduction of the
the overall processing time by limiting the human intervention to costs associated with the study that are sustained by the health-
the quality assessment of the generated model, it would also cause care provider. The pipeline can also be economically viable for os-
an overall increase in the cost of the service. However, in recent teopenic patients, if used in combination with BMD and if the cost
years studies have emerged that present state of the art neural of the simulation service is kept sufficiently low. It is however to
networks that are able to produce high quality segmentation of fe- acknowledge that CT2S is currently not cost-effective when used
murs starting from CT images [35] and generate meshes for finite in clinical studies aimed at the prediction of the risk of fracture
element analysis [36]. Further advancements in the technology will [37]. We recognize that this preliminary health economics analysis
make such tools more abundant in the future and allow an easier is limited to the model established at STH/TUOS, or the UK NHS
integration within the CT2S workflow. system, and needs to be further expanded in order to investigate
To make the workflow sustainable the costs must be balanced the impact of deploying such a workflow at other UK centres and
by the price of the service. An analysis recently reported by Vice- in other countries. Also, more specific economical consideration re-
conti et al. [37] concluded that the CT2S pipeline is economically lated to the use of CT2S in the clinical practice are not possible at
preferable to traditional BMD analysis in clinical studies aiming the moment as currently it is not recognised as a medical device
at strength estimation, which is the main output of the work- by any regulatory authority.
flow. This is based on the observation that CT based FE models Currently, the CT2S service is only available in Sheffield, UK.
require half the subjects needed by BMD methods to achieve the However, besides the meshing and simulation parts and the PIX

7
I. Benemerito, W. Griffiths, J. Allsopp et al. Computer Methods and Programs in Biomedicine 208 (2021) 106200

system, which is internal to STH, the workflow is based on open given time. The workflow can also be used alongside other pre-
source software. The workflow has potential to be deployed in dictive tools, such as FRAX [40], which takes into account patients
other institutions, and The University of Bologna and Istituto Or- medical history, or multiscale probabilistic description of the risk
topedico Rizzoli are currently implementing the CT2S workflow in of falling and the corresponding forces, in order to provide the cur-
Bologna, Italy. A successful multicentre deployment requires that rent absolute risk of fracture, or ARF0 [41]. Work is currently un-
the participating centres possess the necessary infrastructure for derway to extend the mathematical model in order to predict the
transferring the images from the hospital networks to an appropri- risk of fracture in the next 10 years, or ARF10 [42], taking into ac-
ate database, as well as the HPC capability to run the FE analysis. count a statistical model of the bone ageing process for each indi-
HPC systems with the capability of running parallel jobs can guar- vidual. This could be integrated into the current CT2S workflow to
antee that the results are made available to the clinical person- produce a more complete prediction of a patient’s risk of fracture
nel in a timeframe compatible with current clinical workflow, but in the short to medium term.
are not strictly indispensable. As long as calibration data are avail-
able, the workflow is applicable to images acquired with any type 5. Conclusion
of CT scanner. To ensure consistency of the analysis it is fundamen-
tal that the scanning procedures are consistent across the centres The CT2S platform as a proof of concept, is set up and func-
where the workflow is deployed. To this end, different calibration tional. Based on an intuitive and user friendly interface, the CT2S
procedures can be defined according to the CT scanner and acqui- service gives clinical staff access to predictive tools based on en-
sition sequence. Calibration data can be provided either scanning gineering solutions, while avoiding the need for them to undergo
a Mindways phantom (Mindways Inc, Austin, TX) together with extensive training to use the backend of the software. The service
the subject, or retrospectively through an European Spine Phan- has clear potential benefits in assessing the risk of fracture with
tom (ESP). In case an ESP is used, we ensure that the calibra- existing methods, and has the potential to provide clinicians with
tion is uniform across different machines by defining the recom- additional quantitative data for the diagnosis and treatment of os-
mended scanning parameters to be used for the acquisition, in- teoporotic fracture in the future. The thorough description of the
cluding the kernel for image reconstruction. Scanning parameters workflow presented here can foster the wider implementation of
such as voltage, current, slice thickness and pitch are chosen in the workflow in other clinical settings, or its integration within
order to minimise the total radiation exposure as estimated by similar patient-specific model-based systems for support to clini-
the ImPACT dosimetry calculator [38]. However, each healthcare cal decision.
provider adopting the CT2S workflow would need to perform a cal-
ibration study together with the estimation of the radiation expo- Declaration of Competing Interest
sure, in order to produce scans of adequate quality while preserv-
ing the health of the patient. A previous study by Viceconti and The authors declare that they have no known competing finan-
colleagues [37] has shown that, despite the radiation dose associ- cial interests or personal relationship that could have appeared to
ated to a CT2S analysis is higher than the radiation from a BMD influence the work reported in this paper.
study, the patient benefits from an overall reduction of the mor-
tality risk when the additional absorbed radiation is factored to- Acknowledgements
gether with the reduction in death risk from bone fracture. Ma-
chine learning algorithms for scanning optimisation [39] could be This work was supported by the European Union Horizon2020
used, in future developments, for further reduction of the radiation programme (CompBioMed project, grant agreement No 675451;
dose and standardisation of the scanning procedure. CompBioMed2 project, grant agreement No 823712) and the UK
As already mentioned, another service for bone strength es- EPSRC (MultiSim project, grant agreement EP/K03877X/1; Multi-
timation exists. VirtuOst is an FDA-cleared software for bone Sim2 project, grant agreement EP/S032940/1).
strength estimation from medical images, provided as an online
service. It uses as input DICOM images from CT scans. The pipeline References
requires calibration, which can be performed either through a
phantom like in the case of CT2S, or using internal organs and tis- [1] R.L. Winslow, N. Trayanova, D. Geman, M.I. Miller, Computational medicine:
translating models to clinical care, Sci. Transl. Med. 4 (158) (2012).
sues as a reference. Unlike CT2S, whose analysis is limited to the 158rv11-158rv11
proximal femur, VirtuOst can also process images of vertebras. The [2] M. Beier, C. Jansen, G. Mayer, T. Penzel, A. Rodenbeck, S. Siewert, M. Witt,
resulting FE model of the bone, meshed using 10-node cubic ele- J. Wu, D. Krefting, Multicenter data sharing for collaboration in sleep medicine,
Future Gener. Comput. Syst. 67 (2017) 466–480.
ments, is loaded with a single force evenly distributed to the bone [3] Z. Zhang, F.S. Yin, W.K. Wong, N.M. Tan, B.H. Lee, J. Cheng, T.Y. Wong, Ori-
surface by means of a thin layer of plastic [8]. Bone plasticity is ga-light: an online retinal fundus image database for glaucoma analysis and
also taken into account, leading to a nonlinear FE model. CT2S in- research, in: 2010 Conf Proc IEEE Eng Med Biol Soc, 2010, pp. 3065–3068.
[4] M.C. Oliveira, W. Cirne, M.P.M. de Azevedo, Towards applying content-based
stead employs a 8-node tetrahedral discretisation, which results in image retrieval in the clinical routine, Futur. Gener. Comput. Syst. 23 (3) (2007)
smooth meshes, and estimates the bone strength applying a wide 466–474.
array of loads [18]. Material nonlinearities effects are not consid- [5] M.S. Hansen, T.S. Sørensen, Gadgetron: an open source framework for medical
image reconstruction, Mag. Reson. Med. 69 (6) (2013) 1768–1776.
ered in the CT2S workflow.
[6] P.D. Morris, R. Ryan, A.C. Morton, R. Lycett, P.V. Lawford, D.R. Hose, J.P. Gunn,
As mentioned above, CT2S is currently not recognised as a med- Virtual fractional flow reserve from coronary angiography: modeling the signif-
ical device by any regulatory authority. In particular, any medi- icance of coronary lesions: results from the VIRTU-1 (VIRTUal fractional flow
reserve from coronary angiography) study, JACC Cardiovasc. Interv. 6 (2) (2013)
cal diagnosis tools need to be approved by the UK’s Medicines
149–157.
and Healthcare products Regulatory Agency (MHRA), before they [7] Available from: https://heartflow.com Accessed: November 10, 2020.
can be placed on the market for medical purposes in the UK. De- [8] T.M. Keaveny, B.L. Clarke, F. Cosman, E.S. Orwoll, E.S. Siris, S. Khosla, M.L. Boux-
tailed comparison with VirtuOst’s cost model is not possible be- sein, Biomechanical computed tomography analysis (BCT) for clinical assess-
ment of osteoporosis, Osteoporos. Int. (2020) 1–24.
cause of the marked differences in healthcare business model be- [9] E. Banks, G.K. Reeves, V. Beral, A. Balkwill, B. Liu, A. Roddam, Million women
tween the US (largely owned by private companies) and the UK study hip fracture incidence in relation to age, menopausal status, and age at
(publicly funded) menopause: prospective analysis, PLoS Med. 6 (11) (20 09) e10 0 0181.
[10] J. Leal, A.M. Gray, D. Prieto-Alhambra, N.K. Arden, C. Cooper, M.K. Javaid,
Currently the workflow operates with data collected during A. Judge, REFReSH study group impact of hip fracture on hospital care costs: a
one scanning session, and estimates femoral bone strength at that population-based study, Osteoporos. Int. 27 (2) (2016) 549–558.

8
I. Benemerito, W. Griffiths, J. Allsopp et al. Computer Methods and Programs in Biomedicine 208 (2021) 106200

[11] F. Borgström, L. Karlsson, G. Ortsäter, N. Norton, P. Halbout, C. Cooper, [28] Available from: https://docs.microsoft.com/en-us/dotnet/csharp Accessed:
M. Lorentzon, E.V. McCloskey, N.C. Harvey, M.K. Javaid, J.A. Kanis, Fragility frac- November 10, 2020.
tures in europe: burden, management and opportunities, Arch. Osteopor. 15 [29] Available from: https://rabbitmq.com/ Accessed: November 10, 2020.
(2020) 1–21. [30] Available from: https://www.erlang.org Accessed: November 10, 2020.
[12] C.H. Turner, Bone strength: current concepts, Ann. N. Y. Acad. Sci. 1068 (1) [31] P.A. Yushkevich, J. Piven, H.C. Hazlett, R. Gimpel Smith, S. Ho, J.C. Gee,
(2006) 429–446. G. Gerig, User-guided 3D active contour segmentation of anatomical struc-
[13] B.L. Riggs, H.W. Wahner, E. Seeman, K.P. Offord, W.L. Dunn, R.B. Mazess, tures: significantly improved efficiency and reliability, NeuroImage 31 (3)
K.A. Johnson, L.J. Melton, Changes in bone mineral density of the proximal fe- (2006) 1116–1128.
mur and spine with aging: differences between the postmenopausal and senile [32] F. Taddei, E. Schileo, B. Helgason, L. Cristofolini, M. Viceconti, The material
osteoporosis syndromes, J. Clin. Investig. 70 (4) (1982) 716–723. mapping strategy influences the accuracy of CT-based finite element models
[14] P. Ammann, R. Rizzoli, Bone strength and its determinants, Osteoporos. Int. 14 of bones: an evaluation against experimental measurements, Med. Eng. Phys.
(3) (2003) 13–18. 29 (9) (2007) 973–979.
[15] M. Qasim, G. Farinella, J. Zhang, X. Li, L. Yang, R. Eastell, M. Viceconti, Patien- [33] Regulation (EU) 2016/679 of the european parliament and of the council of 27
t-specific finite element estimated femur strength as a predictor of the risk of april 2016 on the protection of natural persons with regard to the processing
hip fracture: the effect of methodological determinants, Osteoporos. Int. 27 (9) of personal data and on the free movement of such data, and repealing direc-
(2016) 2815–2822. tive 95/46/EC (general data protection regulation). OJ l 119, 4.5, 2016. p. 1-88.
[16] Available from: https://www.dicomstandard.org. Accessed: November 10. 2020 [34] D.C. Barber, D.R. Hose, Automatic segmentation of medical images using image
[17] E. Schileo, F. Taddei, L. Cristofolini, M. Viceconti, Subject-specific finite element registration: diagnostic and simulation applications, J. Med. Eng. Technol. 29
models implementing a maximum principal strain criterion are able to es- (2) (2005) 53–63.
timate failure risk and fracture location on human femurs tested in vitro, J. [35] C.M. Deniz, S. Xiang, R. Spencer Hallyburton, A. Welbeck, J.S. Babb, S. Honig,
Biomech. 41 (2) (2008) 356–367. K. Cho, G. Chang, Segmentation of the proximal femur from MR images using
[18] Z. Altai, M. Qasim, X. Li, M. Viceconti, The effect of boundary and loading con- deep convolutional neural networks, Sci. Rep. 8 (1) (2018) 1–14.
ditions on patient classification using finite element predicted risk of fracture, [36] V. Chandran, G. Maquer, T. Gerig, P. Zysset, M. Reyes, Supervised learning for
Clin. Biomech. 68 (2019) 137–143. bone shape and cortical thickness estimation from CT images for finite ele-
[19] Available from: https://ct2s.insigneo.org/static/CT2SApp/multisim/ ment analysis, Med. Image Anal. 52 (2019) 42–55.
CT2S_Service_CT_Scan_Protocol.pdfAccessed: November 10,2020. [37] M. Viceconti, M. Qasim, P. Bhattacharya, X. Li, Are CT-based finite element
[20] W.A. Kalender, D. Felsenberg, H.K. Genant, M. Fischer, J. Dequeker, J. Reeve, model predictions of femoral bone strengthening clinically useful? Curr. Os-
The european spine phantom - a tool for standardization and quality control teoporos. Rep. 16 (3) (2018) 216–223.
in spinal bone mineral measurement by DXA and QCT, Eur. J. Radiol. 20 (2) [38] Available from: http://www.impactscan.org/ctdosimetry.htm Accessed: 10 May,
(1995) 83–92. 2021.
[21] R.E.J. Cooke, M.J. Gaeta, D.M. Kaufman, J.G. Henrici, Picture archiving and com- [39] C.H. McCullough, S. Leng, Use of artificial intelligence in computed tomography
munication system, 2003, U.S. Patent 6,574,629, issued June 3. dose optimisation, Ann. ICRP 49 (1_suppl) (2020) 113–125.
[22] Available from: https://www.sheffield.ac.uk/it-services/research/hpc/ [40] J.A. Kanis, O. Johnell, A. Odén, H. Johansson, E. McCloskey, FRAXTM and the as-
sharc/Accessed: 10 November, 2020. sessment of fracture probability in men and women from the UK, Osteoporos.
[23] D.S. Marcus, T.R. Olsen, M. Ramaratnam, R.L. Buckner, The extensible neu- Int. 19 (4) (2008) 385–397.
roimaging archive toolkit, NeuroInformatics (2007) 11–33. [41] P. Bhattacharya, Z. Altai, M. Qasim, M. Viceconti, A multiscale model to pre-
[24] T.R. Fielding, Architectural styles and the design of network-based software ar- dict current absolute risk of femoral fracture in a postmenopausal population,
chitectures, 20 0 0, Irvine: University of California, Vol. 7. Biomech. Model. Mechanobiol. 18 (2) (2019) 301–318.
[25] Available from: https://www.djangoproject.com/ Accessed: November 10, 2020. [42] E. Siris, P.D. Delmas, Assessment of 10-year absolute fracture risk: a new
[26] Available from: https://www.python.org Accessed: 10 November, 2020. paradigm with worldwide application, Osteoporos. Int. 19 (2008) 383–384.
[27] Available from: https://www.javascript.com/ Accessed: November 10, 2020.