RPL 211

Expert Systems With Applications 244 (2024) 122983
Contents lists available at ScienceDirect
Expert Systems With Applications

journal homepage: www.elsevier.com/locate/eswa
Review
Paradigm shift from Artificial Neural Networks (ANNs) to deep

Convolutional Neural Networks (DCNNs) in the field of medical
image processing
Serdar Abut a, b, *, Hayrettin Okut b, c, K. James Kallail d
a
Department of Computer Engineering, Siirt University, 56100 Siirt, Turkiye
b
Department of Office of Research, University of Kansas School of Medicine, 1010 N. Kansas, Wichita 67214, KS, USA
c
Department of Population Health, University of Kansas School of Medicine, 1010 N. Kansas, Wichita 67214, KS, USA
d
Department of Internal Medicine, University of Kansas School of Medicine, 1010 N. Kansas, Wichita 67214, KS, USA
A R T I C L E I N F O A B S T R A C T
Keywords: Images and other types of unstructural data in the medical domain are rapidly becoming data-intensive.
Deep Convolutional Neural Networks Actionable insights from these complex data present new opportunities but also pose new challenges for clas
Medical Image Processing sification or segmentation of unstructural data sources. Over the years, medical problems have been solved by
Artificial Neural Networks
combining traditional statistical methods with image processing methods. Both the increase in the size of the
Feature Extraction
data and the increase in the resolution are among the factors that shape the ongoing improvements in artificial
intelligence (AI), particularly concerning deep learning (DL) techniques for evaluation of these medical data to
identify, classify, and quantify patterns for clinical needs. At this point, it is important to understand how
Artificial Neural Networks (ANNs), which are an important milestone in interpreting big data, transform into
Deep Convolutional Neural Networks (DCNNs) and to predict where the change will go. We aimed to explain the
needs of these stages in medical image processing through the studies in the literature. At the same time, in
formation is provided about the studies that lead to paradigm shift and try to solve the image related medical
problems by using DCNNs. With the increase in the knowledge of medical doctors on this subject, it will be
possible to look at the solution of new problems in computer science from different perspectives.
1. Introduction accumulation voluminous complex healthcare data contain ample in

formation for precision medicine, and the only way to extract mean
As a sub discipline of Artificial Intelligence (AI), Machine Learning ingful information that will contribute to medical services from big data
(ML) deals with the ability of computers to learn in the same way that is through the effective use of machine learning algorithms (P. Y. Wu
biological creatures learn. Therefore, machine learning uses set of al et al., 2017).
gorithms that parse data, learn from them, and apply what they have The use of computers became widespread after the 1950 s for solving
learned to make a generalization. According to (Mitchell, 1997), a ma problems such as analyzing and modeling large data sets (Kononenko,
chine learning problem consists of three main terms: task (T), perfor 2001). After this period, many techniques were conducted on medical
mance measure (P), and training experience (E). For example, in a data analysis such as regression (Eastwood & Baker, 1983), decision tree
simple medical diagnosis problem, i) task T: recognizing and classifying (Batterham, Christensen, & Mackinnon, 2009; Pauker & Kassirer, 1975),
COVID-19 disease within X-ray images, ii) performance of measure P: is support vector machines (SVM) algorithm (Batra, 2016), naive bayes
percent of true predicted patients against all amount patients, iii) algorithm (Anderson, 1974), k-nearest neighbors algorithm (k-NN) al
training experience E: a database of X-ray images that labels as COVID- gorithm (Binnie et al., 1978) and k-means (Everitt, Gourlay, & Kendell,
19 disease or not. If a computer programs performance improves with 1971).
experience, it can be said that it is learning (Mitchell, 1997). The technological advances for every field of biology, in particular
Rapid advances of high-throughput technologies and fast medicine and healthcare, have generated more structured and
* Corresponding author.
E-mail addresses: [email protected], [email protected] (S. Abut).
https://doi.org/10.1016/j.eswa.2023.122983
Received 24 April 2023; Received in revised form 11 December 2023; Accepted 15 December 2023
Available online 16 December 2023
0957-4174/© 2023 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
S. Abut et al. Expert Systems With Applications 244 (2024) 122983
unstructured data, even to a level where it has become unmanageable adaptively update the data with the intent of learning features specific to
with currently available technologies and conventional analytical ap the data by repeating these operations over all of the hidden layers, with
proaches. Structured health information consists of an easily analyzable each layer learning to detect different features. The power of DCNN
database and carries a range of statistics and classes. The massiveness of comes from a special kind of layer called the convolutional layer. Each
medical data is unstructured information within the variety of numerous layer, except the last one, is followed by a nonlinear activation function.
completely different notes, images, a udio and v ideo recordings, re The function of a DCNN is analogous to the connectivity pattern of nerve
ports, and discharge summaries (Shailaja, Seetharamulu, & Jabbar, cells (neurons) in the human brain, and it is inspired by the organization
2018). The volume and complexity of these data introduce new oppor of the visual cortex.
tunities, but also pose new challenges. This review addresses the importance of changing the paradigm shift
According to the calculations of International Data Corporation from ANNs to DCNNs in the field of medical image processing. Some
(IDC), 270 GB of health data were produced for each person in 2020 limitations of the ANNs and image processing challenges also are dis
(Burghard, 2021). As data grows, the computational methods required cussed. The paper is organized in 8 sections. Section 2 explains the
to cope with the unusual dimensionality become more difficult and purposes image processing approaches are used in the field of medicine
complex. Examples of how the structure of decision trees, which is one of and with what effects these processes are shaped. In section 3, the ne
the classical data analysis methods, has changed from the 1970 s to the cessity of the feature extraction procedure, which is very important in
2000 s are shown in the Fig. 1-A-B. For more detailed information, see image processing and has led to the creation of a large literature, is
publication (P.-Y. Wu et al., 2016). discussed. In section 4, the convolution technique is explained, which is
Rapid advances of high-throughput technologies have changed the an important step in the evolution of ANN models to DCNN models in
execution of health services, and data have become an omnipresent image processing, used in previous studies. In section 5, general infor
concept in our daily lives with the routine collection, storage, process mation is provided about DCNNs and ANNs models. Section 6 is based
ing, and analysis of immense amount of data and changes in the docu on the reasons for the divergence of opinions of scientists in the field of
mentation methods (Pastorino et al., 2019). Current healthcare practices computer vision. Chapter 7 provides information about recent work in
store a large amount of patient data. Health centers store patient in medicine where DCNN models are studied on computed tomography
formation such as diagnosis, medication, laboratory results, and radio images. Section 8 provides information about which types of problems
logical image records, omic* datasets (contain a comprehensive catalog DCNN and ANN models can solve in the medical field. Finally, section 9
of molecular profiles, such as genomic, transcriptomic, epigenomic, provides the concluding remarks.
proteomic, and metabolomic), smart health records, language process The diagram presented in Fig. 3 visually conveys the hierarchical
ing of medical documents, and drug discovery (Jensen, Jensen, & Bru structure that delineates the interconnections and dependencies be
nak, 2012). tween different artificial intelligence (AI) concepts. At the top level,
Health records such as medical diagnostics, medication, and labo “Artificial Intelligence” encompasses the broader field, while “Machine
ratory results are generally lower than 10 KB data for each patient re Learning” resides as a significant subset within it. Machine learning, in
cord. However, the health records that include radiological images are turn, branches into several categories, including “Supervised Learning,”
more than 1000 KB. Hence, low dimensional data could be analyzed “Unsupervised Learning,” “Semi-Supervised Learning,” and “Rein
with classical analysis methods. On the other hand, high dimensional forcement Learning.” The deepest layer of this hierarchy is “Deep
data such as radiologial images need to analyzed with novel machine Learning,” which is a subset of machine learning that uses artificial
learning apporaches (Fig. 2). neural networks with multiple layers (deep neural networks) to auto
Deep learning (DL) is one of the most extensively utilized machine matically learn and extract features from data. The deep learning given
learning approaches in electronic health records. The term ‘‘deep” in DL in Fig. 3 further is divided into specialized domains such as “Convolu
refers to the number of hidden layers. Traditional ANNs have 2–3 hidden tional Neural Networks” and “Feed Forward Neural Networks.” This
layers between input and output layers, while DL networks may have as hierarchical structure visually organizes key AI concepts, providing a
many as tens or hundreds of layers. The DL architectures are used for clear framework for understanding their interconnections and roles
feature selection, classification, dimensionality reduction, or as a sub within the field of artificial intelligence.
module of deeper architecture. They automatically extract and learn
features based on large amounts of data. It essentially consists of a 2. A brief history of image processing and applications in
MLANN (multilayer artificial neural network) and an algorithm that medical science
mimics human neurons, and it automatically processes and learns input
data and passes them to the next layer, which consists of many layers. In Before delving into the intricacies and efficacy of modern medical
these layers, it is possible to deepen the characteristics of the data to be image processing, it is essential to acquire a profound understanding of
learned using multiple layers of this neural network, which result in the fundamental principles governing image formation. In a scientific
deep-learning models with extremely high accuracy, sometimes sur context, an image materializes as a visual representation when incident
passing human recognition accuracy (Matsuzaka & Yashiro, 2022). light interacts with an object and is subsequently detected and recorded
Because our goal in this review is to understand the deep learning by a sensor. Images are obtained with light source and are classified
approach in the medical field, we will focus on the medical image pro according to their light sources (e.g., X-rays, viusal light, gamma rays)
cessing procedure, Deep Convolutional Neural Networks (DCNNs). (Gonzales, 1987). Familiarity with these fundamental principles serves
DCNN is a feed-forward artificial neural network and is comprised up to as the foundation for comprehending the underpinnings of medical
hundreds of layers and each learns to detect different features of an image processing.
image. Filters in the architecture are applied to each training image at When medical image processing was initially employed in the field of
different resolutions, uniquely defining the image and the output of each radiology, the obtained images were quite limited in quality and scope.
convolved image as the input to the next layer. A DCNN consists of: i) an However, with the advancement of technology and the development of
input layer, ii) an output layer, and iii) hidden layers which include this field, medical imaging methods and the acquired images have un
multiple convolutional layers, pooling layers, fully connected layers, dergone significant transformations. This transformation signifies a
and normalization layers. These layers perform operations that major milestone in the progress of medical image processing. In 1895,
Wilhelm Conrad Roentgen discovered X-ray and he produced an X-ray
image of his wife Bertha’s hand (Chodos & Ouellette, 2001). After this
*
Omics data is information generated by studies ending with -omics: geno invention, X-ray imaging became the earliest method used for medical
mics, proteomics, phenomics, etc. diagnostics (Gonzales, 1987).
2
Fig. 1. A demonstration of how models get more complex as data gets bigger. A) An example of basic decision tree approach developed in 1975(Pauker & Kassirer,
1975). B) Decision tree for predicting the risk of major depressive disorder (Image by (Batterham et al., 2009), used under CCBY2.0).
Fig. 2. Electronic health recods and their analysis procedure.
When comparing the first X-ray image taken in 1895 (Fig. 4-A) with a the gray level colors of the pictures are seen (Fig. 4-C,G). The image
sample image taken in 2016 (Fig. 4-D), it becomes clear that imaging processing mechanism is based on the principle of performing mathe
technology has come a long way in terms of resolution. When the win matical operations on each picture by defining it as a numerical matrix,
dows are taken from both pictures to correspond to approximately the as in Fig. 4-C and 4-G.
same area in the real world, it is seen that the detail in one is less (Fig. 4- The period preceding 1986 was marked by foundational research
B) and a more detailed view can be obtained in the other (Fig. 4-E). and technological constraints in medical image processing, with a pri
When both pictures are viewed from a 10x10 window, the pixel values of mary focus on the development of basic algorithms. When examining
3
Box 1
List of Major Abbreviations.
ADMET Absorption, Distribution, Metabolism And ExcretionToxicology.

AI Artificial Intellingence.
ANN Artificial Neural Networks.
CNN Convolution Neural Networks.
CT Computed Tomography.
DL Deep Learning.
DNN Deep Neural Networks.
ML Machine Learning.
MLP Multi Layer Perceptron.
MLR Multiple Linear Regression.
MRI Magnetic Resonance Imaging.
RF Random Forest.
RNN Recurrent Neural Networks.
SVM Support Vector Machines.
Fig. 3. A Hierarchical Representation of Some Artificial Intelligence Concepts.
the studies in the field of medical image processing in the 1970 s, various diagnostic outcomes. Saito and Hayashi (1982) tackled the challenge of
methods and solutions were presented. (Sutou & Kahara, 1975) exam solving the inverse scattering problem in the context of medical image
ined and discussed feature extraction methods for image processing and processing, with a specific emphasis on its application to ultrasonic
medical image processing applications. Furthermore, Altschuler and fields. Udupa, Srihari, and Herman (1982) pioneered the development
colleagues (1980) pioneered the development of a software package of algorithms designed to detect boundaries in multidimensional spaces,
designed to reconstruct the dynamic structure of the human heart using a critical component for the processing of time-varying images and
cone beam X-ray projections. In a study conducted by Semmlow, Hand, computerized tomography data. In his work from 1982, Herman
Ackerman, and Alcorn (1980), significant attention was given to the explored computer-based methodologies for the reconstruction of the
design and appropriateness of a computer-controlled apparatus meant dynamic motion of the human heart from X-ray projections, shedding
for the automated retrieval of medical images. Additionally, research light on the intricate process of capturing a beating heart in digital form.
carried out by de Valk, Eijkman, and Vintcent (1981) focused on psy In 1986, Lerallut, Gambier, and Fontenier brought forth the concept of
chophysical metrics pertinent to medical image processing, encom semi-automatic systems for medical image processing, along with a
passing parameters such as the detection index, likelihood ratio, and critical examination of their inherent limitations. Around the same time,
4
Box 2
. Evolution of Artificial Intelligence, Machine Learning and Deep Learning.
Artificial Intelligence (AI): A broad scientific discipline with its roots in philosophy, mathematics and computer science that aims to understand
and develop systems that display properties of intelligence.
Machine Learning (ML): A sub discipline of AI, where computers programs (algorithms) learn associations of predictive power from examples in
data. Machine learning is most simply the application of statistical models to data using computers. Machine learning uses a broader set of
statistical techniques than those typically used in medicine.
Deep Learning (DL): A sub-discipline of ML and based on models with less assumptions about the underlying data, therefore, are able to handle
more complex data. Deep Learning Deep learning methods allow a machine to be fed with large quantities of raw data and to discover the
representations necessary for detection or classification. Deep learning methods rely on multiple layers of representation of the data with
successive transformations that amplify aspects of the input that are important for discrimination and suppress irrelevant variations. Deep
learning may be supervised or unsupervised.
Supervised Learning (SL): Training computer programs to learn associations between inputs and outputs in data through analysis of outputs of
interest defined by a (typically human) supervisor. Once associations have been learned based on existing data they can be used to predict future
examples. This is one of the most established areas of machine learning with multiple examples inside and outside health care.
Unsupervised Learning: Algorithms in this category use machine learning algorithms to analyze and cluster unlabeled datasets. These algorithms
do not need human intervention when revealing hidden patterns or meaningful data components in the data (Corp., 2020). The ability to detect
similar and divergent regions in data makes these algorithms an ideal tool for segmentation purposes in medical image processing.
Convolution: It is the process of obtaining a third new matrix with mathematical operations applied to two matrices. The feature that discriminate
the convolution operation from other mathematical operations is that the second matrix is relatively smaller than the first matrix and the second
matrix is subjected to mathematical operation with all its elements by convolute over the first matrix. In the convolution process under the
subject of image processing, the first matrix is called an image, the second matrix is called a filter or kernel matrix.
Fig. 4. Development stages of X-ray imaging. A- The first x-ray image discovered by Roentgen in 1895 (Chodos & Ouellette, 2001) and marking a window of 50x50
pixels on the picture. B- The appearance of the window marked in the image A and the marking of another window with a size of 10x10 pixels. C- The appearance and
pixel values of the 10x10 window. D- An example of modern x-ray images(Image by (Stachowiak, Wolski, Woloszynski, Podsiadlo, & Biotribology, 2016), used under
CCBY 4.0) and marking a window of 200x200 pixels on the picture. E-F-G- appearances of windows obtained in 200x200, 50x50 and 10x10 sizes, respectively.
Stansfield took a pioneering step by conceiving and putting into practice observed, along with the advancement of telemedical applications. In
ANGY, an expert system meticulously crafted for the automated seg (Michael & Nelson, 1989), a model-based computer vision system called
mentation of coronary vessels from digital subtracted angiograms. HANDX was designed and implemented to automatically separate bones
In subsequent years, medical image processing studies have shown from digital hand radiographs. (Inglis, Windolf, & Pannike, 1990)
the development of more advanced 3D imaging, data analysis, and the focused on medical image processing for non-invasive digital bone
use of computer technologies for more effective and detailed image density measurement using a personal computer. (Akisada, 1990) ex
processing methods. Additionally, an increase in the integration be plains the activities of the Medical Image Processing System (MIPS)
tween medical imaging and medical information management has been plan, a standard plan for online communication of image data in Japan.
5
(Dai, 1991) focused on clinical applications in China using the “IA-87

Medical Image Processing System” for quantitative analysis of left ven
triculograms and coronary arteriograms, providing normal values for
this method. In (Appelman, Zuiderveld, Gash, & Viergever, 1993), a
medical image input/output library based on data dictionaries was
presented, addressing image input/output issues, particularly in the
context of medical image processing. (Noz et al., 1993) describes the
development of a graphical interface that enables the manipulation of
medical image processing data with “click and select” ease. (Lemke,
1993) focused on projects in Europe where prototype systems for med
ical image processing and management were developed, based on
modern computer and communication technologies. (Wiegand, Page, &
Channin, 1993) discussed the development of medical image processing
during the era of high-performance computing, addressing techniques
that enable the integration of multiple imaging modalities, flexible
analysis environments, and easy transmission of PACS (Picture
Archiving and Communication System). (Udupa & Goncalves, 1993) Fig. 5. Google Ngram results of “feature extraction” and “dimensionality
addressed the importance of three-dimensional (3D) visualization in the reduction” terms versus years (Michel et al., 2011).
medical field and presented a framework that combines different 3D
imaging methods. (Udupa & Goncalves, 1993) discussed the creation of to automate tongue diagnosis. (Gao et al., 2023) introduced U-MLP, a
large medical image databases and their association with medical in MLP-based ultralight refinement network, for medical image segmen
formation, addressing challenges in this field. (Odaka, Takahama, tation. U-MLP aimed to achieve high accuracy with fewer computational
Wagatsuma, Shimada, & Ogura, 1994) described the development of a and parametric resources compared to CNN and Transformer models.
visual data analysis system used to manage large amounts of medical (Monroy et al., 2023) presented an automated framework for assessing
image data, featuring a user-friendly graphical interface. and tracking chronic wounds using deep learning. This framework
Huda and Abraham (Huda & Abrahams, 2015) further clarified the worked with RGB images captured with smartphones and included
concept of resolution in radiological imaging in their studies. They used wound detection, segmentation, and quantitative analysis. (B. Wu et al.,
the words blur and sharpness to express the concept of resolution more 2023) addressed the removal of rigid motion artifacts in MRI using an
clearly. With the illustrations they created, they explained in detail how unsupervised dual-domain disentangled network named UDDN. This
the resolution is measured in the field of radiological imaging, the approach effectively removed unwanted artifacts from images and
concept of pixel size and the factors affecting resolution. Digital radio demonstrated promising results. (F. Zhang, Wang, Zhao, & Zhang, 2023)
graphs usually are composed of a 2000x2500 numeric matrix (Huda & explored the applications of Generative Adversarial Networks (GANs) in
Abrahams, 2015). A digital mammogram consists of a numeric matrix of medicine and emphasized their significance in clinical medical research.
approximately 3000x4000 and consists of 12 million pixels (Samei, GANs were found to have potential in privacy protection and medical
2005). Huda and Abraham evaluated how many line pairs per milli diagnosis. (Sun et al., 2023) evaluated the impact of a novel denoising
meter of an image is formed in terms of different imaging systems. These algorithm on image quality in dual-energy abdominal CT scans of obese
imaging systems are Computed Tomography (CT), Fluoroscopy, Digital patients. The test algorithm demonstrated better image quality and was
angiography, Radiography and Mammography, and they stated that the preferred by readers. (Yuan, Zhou, Guo, Li, & Yu, 2023) introduced
images obtained with these imaging systems consist of approximately HCformer, a hybrid CNN-Transformer network for LDCT image
0.7, 1, 2, 3 and 7 line pairs per milimeter, respectively (Huda & Abra denoising. HCformer combined the strengths of CNN and Transformer
hams, 2015). structures, improving image quality while reducing noise. (Schwartz
The human eye can perceive about 5 line pairs per millimeter when et al., 2023) evaluated a denoising algorithm’s impact on dual-energy
looking at an image at a distance of 25 cm to 10 in. (Samei, 2005) and abdominal CT imaging of obese patients. The test algorithm out
the discernable line pairs are halved as the distance is doubled. In this performed existing methods and was preferred by readers.
case, for instance, the mammography images contain so much detail that
the human eye cannot easily perceive them without any assistive tool. 3. Feature extraction in medical image processing
The main idea behind image processing is that it filters out the important
region or features with a lower dimension from images that have a high Feature extraction is a crucial operation for the dimensionality
dimension. In this way, it could be possible that the human eyes analyze reduction process, in which the raw and big data are divided and
images effectively without any assistive tool. In Fig. 5, the results of reduced to more manageable components, thus, effectively reducing the
Ngram Corpus (Michel et al., 2011) can be seen and the interest towards amount of data. Therefore, effectively reducing the big data, the reduced
the “feature extraction” term which is subtopic of “dimensionality features are easy to process, but still able to represent and describe the
reduction”, has increased in a distinctive way. real data set with precision and originality. The development of the
While previous studies have shown limitations in computer tech image processing field has long been dependent on the development of
nology for medical image processing and management, in recent years, feature extraction methods, until deep convolution neural networks are
thanks to advancing computer and communication technologies, the used.
digitalization and transmission of medical images have become more Biomedical signal data, such as magnetic resonance imaging (MRI),
effective. In this evolving process, the use of digital imaging methods electroencephalogram (EEG), gait rhythm (GR), electrooculogram
such as digital radiography has become widespread, providing greater (EOG), electrocardiogram (ECG), X-rays, computed tomography (CT)
opportunities for more precise diagnoses and surgical planning in pa scan, magnetic resonance imaging (MRI), and nuclear medical imaging
tient care. Developments in the modular structure and standard in have high dimensionality. For analyzing these data, dimensionality
terfaces of medical image processing software have helped create an needs to be reduced. Dimensionality reduction is the process of trans
infrastructure that can better adapt to the specific needs of clinical de forming high-dimension data to low-dimension data. But, meaningful
partments. In (H. Xu, Chen, Qian, & Li, 2023), a two-stage segmentation information should be preserved while this process is conducted (Van
method for sublingual veins in Traditional Chinese Medicine images was Der Maaten, Postma, & Herik, J.V., 2009). Therefore, dimension
proposed. This approach utilized compact fully convolutional networks reduction can be performed assuming that some important parts of the
6
data are related directly to the problem. It is of great importance to be During image processing, the coordinates of some critical points on
able to represent the data with as few underlying parameters as possible. the image must be found to make calculations such as distance from the
These parameters are called intrinsic or effective dimensionality of the data image, area, growth rate, and volume increase. The properties needed to
set (Fukunaga, 2013). make such a calculation are called geometric properties. At the same
Feature extraction methods have been developed to obtain infor time, features that require the evaluation of many parts of the image
mation about the useful parts of the image in accordance with the image together are often needed in medical image processing. Such features are
processing process to be executed, and to get rid of the useless parts of called spatial features and these features are used for texture analysis
the image. Feature extraction approaches focus primarily on edge problems. For example, the algorithm of Lei et al.(Lei, Zhao, Wang, Yu,
detection processes (Canny, 1986; Sobel & Feldman, 1973). & Guo, 2017), which they developed to provide early diagnosis of
During the processing of microarray images, which is one of the cirrhosis with machine learning, provides a solution to such a texture
stages of gene expression analysis, areas that should not be included in analysis problem (Fig. 6).
the background sometimes appear as the background on complementary
DNA (cDNA) microarray and spots, or vice versa. Kim et al. have solved 4. What is image Convolution?
this problem with an approach based on the edge detection algorithm
(Kim, Kim, & Lee, 2001). The concept convolution is frequently used in image processing for
Williams et al. (Williams et al., 1989) developed a edge detection both feature extraction and image filtering. It is also a simple mathe
algorithm to calculate tumor volumes in the liver. Liver CT images matical operation of passing a signal through a linear process which is
containing 19 focal tumors were used to determine the performance of fundamental to many common image processing operators. The
the method they developed. The cross-sectional areas of the CT images convolution process has been used in the field of image processing for a
calculated as a result of the edge extraction algorithm are compared long time, but it has taken a place in deep learning as a process that
with the results of an expert CT interpreter analyzing the same images. provides very effective solutions in deep convolutional neural networks
Edge detection is mostly used for purposes such as area length and (DCNNs) structures, which is a popular topic of recent times. A simple
approximate volume calculation on a medical image and an edge understanding of this process is essential to understanding DCNNs
extraction algorithm was presented by Hu et al. (Hu, Rogers, Coast, structures.
Kramer, & Reichek, 1998), for solving problems such as fuzzy boundries, In Fig. 7 the image consists of pixels and a numeric matrix. Convo
inconsistent image contrast, missing edge, and complicated background. lution is based on creating a new image by performing some mathe
There are many feature extraction methods to represent the image with matical operations by convolution of a kernel on a digital image as
less data. For example, Fourier features have been used in medical image illustrated in Fig. 7. For example, the kernel matrix used as the edge
processing, early detection of microcalcifications (Kothapalli, Yell removal filter proposed by Irvin Sobel in 1973 (Sobel & Feldman, 1973)
eswarapu, Naraharisetty, Wu, & Rao, 2005), and differential diagnosis for the x horizontal component is shown in Fig. 7.
of liver tumor on computed tomography images (Arakeri & Reddy, In Fig. 7, the kernel matrix is applied to a 10x10 numeric matrix
2012). Gray Level Co-occurrence Matrix (GLCM), was used for feature which is a part of a large image. Kernel is an auxiliary matrix that
extraction in the model created for the classification of kidney ultra operates on the original image and is used to slide over the original
sound images (Priyanka & Kumar, 2020)and used in a study comparing image. Valuable information is retrieved from the convolutional matrix
the models to be used to remove noises in medical images (Beksaç, by the kernel applications. A kernel operation starts at the top left corner
Egemen, Izzetoglu, Ergün, & Erkmen, 1996; Raj, Vijayalakshmi, & Priya, of the image and then slides on each part of the pixel sequentially. The
2019). original image consists of x(row,column) pixels, and the resulting image
Fig. 6. The spatial properties of the images (Lei et al., 2017) A- Original image of normal liver and patterns B- Original image of cirhosis and patterns C- Patterns of
normal livers D- Patterns of cirhosis and results of alghorithm.
7
Fig. 7. Convolution of 3x3 kernel matrix to a 10x10 image matrix.
∑∑
consists of y(row,column) pixels. The pixel with the value of 77 at x(2,2) be represented mathematically as: F(x, y) = i j f(x + i − 2, y + j − 2)
position is overlapped with the middle part of the pixel. The sum of the h(i, j). Here, F(x,y) is a pixel of the output image, f(x,y) is a pixel of
product of the pixel values of the original picture and the corresponding original image and h(i,j) is the kernel where i,j∊{1, 2, 3}. This operation,
values in the kernel is calculated. The resulting value 88 creates the y(2,2) applied for the value of 77, is applied to the pixels from x(2,3) to x(2,9)
value of the result image. The convolution operation applied above can (blue rectangle), and the resulting pixel values from y(2,3) to y(2,9) (green
Fig. 8. Convolution operation of the 3x3 kernel matrix proposed in Fig. 3 on an original image A- An original image B- Convolutional results of 3x3 kernel matrix.
8
rectangle) are obtained. When this process is applied for each pixel of scattered and in small pieces, but partially pushes the regions that are in
the original image, new values of the resulting image are created. one piece and large pieces into the background. This procedure increases
Making a complete cycle in this way is called convolution. the visual density of microcalcifications and decreases the densities of
Fig. 8 shows the application of the convolution of a 3x3 kernel matrix other large tissue masses. In this way, microcalcifications that were not
used by Sobel (Sobel & Feldman, 1973) to an X-ray image. Convolution visually noticeable before the use of image processing becomes more
can also be applied with larger kernel matrices. The size and numerical visible.
values of the kernel matrix are determined by the nature of the study. In a study where the Computer Aided Diagnosis (CAD) method based
In a study (Wilson, Hargrave, Mitra, Shieh, & Roberson, 1998) to on patch-based feature extraction was proposed (Nishio & Nagashima,
detect microcalcifications early, the convolution technique was used in 2017), the discriminative properties of the lung nodule were extracted
the image enhancement process, which can solve the problems caused by combining the image convolution technique with principal compo
by low contrast. In the study, the places where calcium deposits occur nent analysis (PCA) and pooling operation. Seventy-three lung nodules
were sought to be detected by combining image convolution with other revealed on 60 sets of computed tomography (CT) images were
image processing techniques such as histogram manipulation and sta analyzed. In their study, PCA was used to determine the numerical
tistical filtering. Because microcalcifications are usually very small, they values of the kernel matrix that will solve the problem most effectively.
have a high spatial frequency. For this reason, a kernel matrix that acts In this unique study, the extracted features were classified by SVM and
as a high-frequency filter. It highlights the image regions that are the area under the ROC curve of the system was reported as 0.83. To use
Fig. 9. Representation of 16-depth convolution layer formation using the convolution principle in DCNNs models.
9
CT density heterogeneity as the nodule feature, the proposed method Having a large number of parameters makes it easy to obtain pre
used image convolution and pooling operations. dictive values close to the target values of the ANN model, but this
In Fig. 9, 16 different matrices obtained by convoluting 16 different causes another problem called overfitting. Overfitting is unavoidable
filters with the lung X-ray image in the input are displayed. In Fig. 9, a without practicing any regularization if the number of observations in
representative deep Convolutional Neural Network (CNN) architecture the training data set is less than the number of network parameters
is seen, demonstrating a model that makes decisions for the output layer (weights and biases). In the case of overfitting, the outputs deviate
as ’Disease’ and ’Normal’. The convolution layer is characterized by drastically from the target values when examples that the system never
dimensions of 222x222x16, indicating height, width, and depth. As encountered during training are given as inputs (Gianola, Okut, Weigel,
observed in Fig. 7, the output image obtained depends on the applica & Rosa, 2011). There are studies that use regularization techniques to
tion of the convolution technique and can be of smaller dimensions solve the overfitting problem in ANN models (Campos et al., 2010;
compared to the original input image. While the convolution technique Gianola et al., 2011; H. Okut, 2016; H. Okut, Gianola, Rosa, & Weigel,
has been used in image processing for a long time, it has become popular 2011; H. Okut et al., 2013). Regularization is the technique of letting
in modern Deep CNNs due to factors such as the increase in processing parameter bias in the direction of what are considered to be more
power and the availability of large datasets that allow the discrimination probable values, which decreases the variance of the estimates at the
of disease patterns from non-disease patterns. The paradigm shift cost of presenting bias (H. Okut, 2016).
created by these two factors is discussed in the next section. On the other hand, compared to ANNs with layers of similar di
mensions, Deep Convolution Neural Networks (DCNNs) have much less
5. Artificial Neural networks (ANNs) and deep Convolutional local-connectivity and parameters, making it easier to learn the re
Neural networks (DCNNs) lationships between inputs and outputs (Krizhevsky, Sutskever, & Hin
ton, 2017). Having a locally connection structure provides great
Artificial neural networks (ANNs) mimic the function of the human advantages to DCNNs in problems where important but little informa
brain and they are capable to implement massively parallel computa tion is wanted to be extracted from big data.
tions. The ANNs can capture the highly nonlinear associations between Fig. 8 illustrates the rough functioning of ANN and DCNN based
predictor variables and outcome for conduct mapping, function models for the machine learning classification task used in medical
approximation, classification, and pattern recognition processing. As a image processing. In Fig. 10, it is assumed that machine learning systems
universal approximator, they are capable to optimize the learning al use medical images as inputs and predict that they classify disease or
gorithm adaptively to learn the complex functional forms. non-disease as output. As can be seen, ANN models are fully connected
ANNs are designed as input (predictors), hidden layer and output and need a dimensionality reduction process before themselves. But
(outcome) layer. The number of neurons in the input layer depends on DCNN is locally connected and does not need any feature extraction or
the size of the input data. The output layer is designed according to the other dimensionality reduction process. DCNN structures can take pic
objectives of ANN such as regression and classification. Depending of the tures as inputs as they are.
complexity of the problem, the number of hidden layers and the number Other than ANNs and DCNNs, many other machine learning algo
of neurons in these layers can be increased. As depicted in upper right rithms have been applied for classification of diseases from medical
panel of Fig. 10, the neurons included in the each layer of the ANNs are health records. No single method performs best in all cases when efforts
fully connected. That is, each of the neurons in the layers is connected to comparing the predictive ability of linear and non-linear approaches for
each of the neurons in the consecutive layer. Each connection is called the classification of complex diseases. The predictive ability of non-
weight or parameter (H. Okut, 2016). Each of these weights is calculated linear as well as linear algorithms depend on many issues including
at each iteration of the training stage of the neural networks. the number of individuals in the training data set, the contribution
Fig. 10. General diagram of models based on ANN (right top panel) and DCNN (right bottom panel) methods for classification problem in medical image processing.
10
demographic and clinical background, the genetic architecture of dis

eases, and effective population size. Table 1 depicts some studies
designed to deal with the classification of diseases in the field of
medicine.
6. A paradigm shift from ANNs to DCNNs
Fig. 11 is provided to visually demonstrate the differentiation in how

an image is used as input for an Artificial Neural Network (ANN) model
compared to a Convolutional Neural Network (CNN or DCNN). As
depicted in the figure, when providing an image to a simple artificial
neural network, the network treats each pixel or feature in the image as
an independent input. This implies the necessity to process a substantial
volume of input data, making it challenging to capture relationships
between features in this manner. In contrast, DCNN architectures are
capable of hierarchically extracting features from images by employing
convolutional operations with filters and kernels. Consequently, DCNN
models aid in automatically identifying edges, shapes, and other sig
nificant features within the image. DCNN networks possess the ability to
learn and extract these features automatically.
ANNs are fully connected, containing a large number of parameters,
with each pixel in the input image having a weight associated with every
neuron. This leads to ANNs requiring more extensive datasets for
training. Conversely, DCNN architectures, through their locally con Fig. 11. Illustration of the most fundamental differences between ANN and
nected structures, focus on specific regions of the input image. Conse DCNN architectures.
quently, DCNN models require fewer parameters while addressing the
same problem. These advantages have positioned DCNNs as popular and problem. They thought that image analysis could not be done only by
efficient tools in the field of medical image processing, offering compact inputting pictures into the neural network and providing the labels to
solutions. the output of the neural network. Information specific to the nature of
In the early 2010 s, the digital image specialists in the field of the problem had to be given to the network. If such a global solution
computer vision had a narrow window in which ANNs could analyze an really existed, it would be able to solve every classification problem and
image with or without feature extraction. According to them, an image it would be unnecessary to create separate designs for each problem. In
processing procedure had to include specific feature extraction methods. such a conjuncture, a paper by Yann LeCun and his collaborators was
Nearly all experts thought every image analysis problem must have a rejected by a prestigious computer vision conference as it used purely
feature extraction method which characterized the nature of this neural networks and did not contain any hand-designed designs suited
Table 1
- Some studies using ANN as a classifier in solving medical problems.
Study Problem Dataset Description Feature Extraction Classifier Accuracy
methods
(Shuqi, Hong, Classification of normal, 322 mammogram images. Scale Invariant Feature BP Neural Network 96.57 %
Zheng, & benign and malignancies (208 normal, 208 benign or Transform (Lowe, 1999)
Chaomin, 2017) malignancies)
(Grossi, Olivieri, & Autism spectrum disorder EEG signals of 15 ASD Multi scale Entropy Sine Net Neural Networks, Logistic Function, 84 % − 92.8 %
Buscema, 2017) (ASD) classifier subjects and 10 normal Sequential Minimal Optimization, K- (Leave One
subjects Contractive Map (K-CM), Naïve Bayes, Out)
Random Forest
(Rajesh & Malar, 1- prediction whether there 20 brain MRI images (10 Rught Set Teory ( Feed Forward Neural Network 95 %
2013) is a tumor. with tumor, 10 without Pawlak, 1982)
2- if there is a tumor, to tumor)
predict whether it is benign
or malignant
(Kurnaz, Dokur, & Tissue segmentation Ultrasound images Discrete Fourier Incremental neural network 98 %
Olmez, 2007) transform. Discrete
cosine transform
(Chak, Navadiya, Kidney stone classification 100 kidney images (normal Gray level Co- Support Vector Machines 99 %
Parikh, & Pathak, and abnormal) occurrence Matrix
2019) (GLCM)
(Maroulis, Colorectal cancer diagnosis Normal and abnormal Discrete Wavelet Multilayer perceptron neural network 95 %
Iakovidis, and detection of pre- endoscopy images Transform, second order
Karkanis, & cancerous polyps statistical features
Karras, 2003)
(Dudhane & Talbar, Interstitial lung disease Interstitial lung diseases Multi-Scale Directional Feed-forward neural network 90.44 %
2018) (ILD) pattern classification database (Depeursinge Mask Pattern (MSDMP)
et al., 2012),
VIA/ELCAP CT Lung Image
Dataset
(Liu, Du, Wang, Diagnosis of premature ECG Records of MIT-BIH Derivatives of Lyapunov Learning vector quantization (LVQ) neural 90.26 % −
Zhou, & Zhang, ventricular contraction arrhythmia database. ( exponents network 98.90 %
2015) Moody & Mark, 2001)
11
to the nature of the problem the vision system was solving. The com 7. Revolutionizing medical computed tomography through
puter vision community showed resistance in appreciating a general DCNNs
learning procedure that could be a common solution to most problems.
The critical threshold for using this general-purpose approach to replace Medical image processing has become an indispensable tool in
complex programming procedures was exceeded by the idea that there modern medicine. Processes critical to the field, such as disease diag
had to be enough computation and enough data. The deep neural nosis(Safdarian & Dabanloo, 2021; J. Xu, Ren, Cai, & Zhang, 2023),
network called SuperVision developed at the University of Toronto has treatment planning (Islam, Wijewickrema, & O’Leary, 2022; Palkovics,
opened the door to a paradigm shift that might have happened earlier, Mangano, Nagy, & Windisch, 2020; Sinzinger, Astaraki, Smedby, &
by halving the error rate on the object recognition problem from images. Moreno, 2022), surgical guidance(Rao, K, Singh, & J, 2022), and patient
The effectiveness of CNNs has been demonstrated on the ImageNet monitoring(Heckel et al., 2014), rely on imaging technologies. In this
(Deng et al., 2009) dataset containing approximately 15 million labeled context, computed tomography (CT) scans form one of the cornerstones
high-resolution images divided into 22,000 categories (Krizhevsky et al., of the medical world, producing high-resolution 2D and 3D images that
2017). allow for detailed examination of internal organs.
The sparsity connection principle in the convolution operation re Nevertheless, conventional approaches within the domain of medical
duces computational complexity (Amer & Maul, 2019) and overfitting image processing exhibit inherent constraints. Challenges such as the
(Hayrettin Okut, 2021) while maintaining information transfer effi need for manual processing, the ever-increasing volume of image data,
ciency. The convolution and pooling operations together serve to extract dependence on expert skills, and errors in the diagnostic process are
location invariant features from images (Lamba, Baliyan, & Kukreja, significant issues that need to be addressed in the field of medical image
2022)(Fig. 12). The ability to create many feature maps using the pixels processing. In this context, deep learning methods, particularly Deep
of the same image is one of the advantages of the convolution process’s Convolutional Neural Networks (DCNNs), have brought about a revo
ability to share parameters. lutionary change in the field of medical image processing. DCNNs have
Using nonlinear rectifiers is the best way to improve the performance surpassed the limitations of traditional methods due to their ability to
of object recognition (Jarrett, Kavukcuoglu, Ranzato, & LeCun, 2009). A process large datasets and their automatic learning capabilities. This
handicap of common sigmoid functions such as arctangent, hyperbolic provides a range of advantages, such as increasing diagnostic accuracy
tangent, and logisitic function, which are used as rectifiers in ANNs (Rompianesi, Pegoraro, Ceresa, Montalti, & Troisi, 2022), expediting the
structure, is that the derivative of the function approaches zero where diagnostic process (Liang et al., 2020), and enabling in-depth analyses
the values go to extreme points, and learning slows down excessively that were previously not possible.
(vanishing gradient). Therefore, there is a need for rectifiers that are In (Ardila et al., 2019), an artificial intelligence model utilized to
actually nonlinear but similar in structure to linear functions. Rectified enhance the accuracy and consistency of lung cancer screening has
Linear Unit (ReLU) is such a rectifier (Fig. 12). It is a function with low yielded successful results with an AUC (area under the curve) perfor
computational cost when its derivative is taken. With these features, mance of 94.4 % in 6,716 cases. This model is acknowledged for its
ReLu is an efficient nonlinear function used in deep networks (Good capacity to optimize the screening process, leading to heightened global
fellow, Bengio, & Courville, 2016). acceptance and improved accuracy in lung cancer screening.
The pooling layer is smaller representations of the input image. In this In (Lee et al., 2022), an artificial intelligence model named ADA was
way, when a region in the input layer is removed, most representations developed to analyze brain computed tomography (CT) images of
in the pool layer are preserved (Goodfellow et al., 2016). This layer healthy individuals and prioritize brain CT images by detecting emer
reduces the dependency of the results that the model will produce on the gency findings. ADA demonstrated a success rate of 0.85 (0.81–0.89)
entire input (Fig. 12). and 0.87 (0.85–0.89) in detecting emergency cases. Additionally, the
In a classification problem, fully connected layer calculates the implementation of ADA led to a noteworthy reduction in post-triage
probability of each label. In a regression problem, it calculates the waiting times and radiology report turnaround times.
estimated output values on the last data reduced to a single dimension. It has been determined that cone-beam computed tomography
This part contains all the advantages of ANNs models (Fig. 12). (CBCT) images synthesized using magnetic resonance imaging (MRI)
exhibit fewer artifacts and noise compared to original CBCT images,
enabling clearer visualization of hard tissues (Choi et al., 2023). As a
consequence, the successful implementation of radiation-free imaging
techniques has proven to be a viable and effective alternative to
Fig. 12. Fundemental structures and principles that operate in CNN models.
12
traditional CBCT, marking a significant advancement in medical imag prognosis prediction. The integration of features extracted from chest CT
ing technology. volumes with clinical data enabled a successful risk stratification of
Another implamentation of artificial intelligence (Jaskari et al., COVID-19 patients.
2020), demonstrated that an artificial intelligence model used in dental
implantology shows high success in the automatic localization of 8. Problems that ANNs and DCNNs offer solutions
mandibular canals. The model was trained on 637 CT volumes with
diverse clinical cases and evaluated on 15 volumes with accurately an In Table 1 and Table 2, studies designed to be a solution to some
notated mandibular canal levels. The results obtained illustrate the problems encountered in the field of medicine are given. These studies
significant potential of this deep learning model to greatly reduce the have been selected to reveal the difference between the use of ANN and
requirement for manual labor.in the workflow of dental implantology. DCNN structures. If Table 1 and Fig. 8 are taken together, machine
(Zeleznik et al., 2021) has achieved the capability of accurately pre learning models using ANN structures need a dimensionality reduction
dicting cardiovascular events by utilizing a deep learning system phase. On the other hand, DCNN structures do not need such a phase
capable of automatically measuring coronary artery calcium. with the advantage of being locally connected.
Strong correlations were obtained using an artificial intelligence There are some valuable review articles about unsafe CNNs in the
model utilized to predict body weight from computerized tomography medical field as diagniosis of COVID-19 (Soomro et al., 2022), counting
(CT) images, achieving a clinically acceptable accuracy in body weight microorganisms (Ma et al., 2022; J. Zhang et al., 2022), image
prediction (Ichikawa, Hamada, & Sugimori, 2021),. As a result of the captioning (Beddiar, Oussalah, & Seppanen, 2022), medical instrument
(Soffer et al., 2021), it has been determined that deep learning algo detection on ultrasound images (H. X. Yang, Shan, Kolen, & de With,
rithms are successful in detecting pulmonary embolism (PE) in 2022), breast cancer classifications (Murtaza et al., 2020), image anal
Computed Tomographic Pulmonary Angiography (CTPA) images and ysis on CT and PET images (Domingues et al., 2020) and transfer
they yield an acceptable number of false positive results. learning in medical image processing (Atasever, Azginoglu, Terzi, &
Shi et al. (2020) demonstrated the efficacy of a deep learning model, Terzi, 2023).
trained on 1,177 bone-removed computed tomography angiography As an example of recent studies, studies aiming to automatically
cases validated by digital subtraction angiography, for the diagnosis of detect mandibular canal using CNNs (Kwak et al., 2020; Vinayaha
intracranial aneurysms. In another study, Chao et al. (2021) utilized a lingam, Xi, Berge, Maal, & de Jong, 2019) have been conducted. The
deep learning model to prognosticate the mortality risk associated with position of this canal must be determined correctly to prevent nerve
cardiovascular disease (CVD) in patients diagnosed with cancer, injury during surgical procedures such as implant installation, third
demonstrating a noteworthy degree of success in the domain of risk molar extraction, and various other craniofacial surgeries including
assessment. orthognathic surgery. In addition, it is important to accurately detect the
ResNet-34 is a distinct iteration of the Residual Neural Network position of this canal in the diagnosis of lesions adjacent to the
(ResNet) structure, incorporating a total of 34 layers to mitigate the mandibular canal and in the planning of oral and maxillofacial surgeries.
challenge of vanishing gradients in deep neural networks. In a recent In a study (Vinayahalingam et al., 2019) in which the mandipular
study conducted by Nomura et al. (2023), the ResNet-34 model, trained canal was detected automatically; the success rate was reported as 84.7
on images featuring oil, exhibited the most remarkable performance in %. In this study, images obtained by using the orthopantomogram
the diagnosis of lymphedema. The performance of tooth segmentation in (panaromic image) imaging technique were used, which have higher
cone-beam computed tomography images has been successfully resolution than CBCT but cause more radiation exposure.
improved using the 3.5Dv5 U-Net model in (Hsu et al., 2022). In (X. Later, Kwak et al. designed a study (Kwak et al., 2020) that aimed to
Yang et al., 2023), a technique was developed for the automated pre automatically locate the mandibular canal on CBCT images, which
diction of right lobe graft weight using deep learning algorithms, would further facilitate clinical use. Some pre-processing procedures
resulting in a high level of success. were made on the Cone beam computed tomography (CBCT) of 102
In the study by Heo et al. (2022), findings suggest that an artificial patients, and the upper jaw region that was not related to the problem
intelligence (AI)-based predictive model for traumatic intracranial was removed from the pictures, and the mandibular part was obtained.
hemorrhage effectively influences the clinical decision-making process This process has been done in order to increase the performance of CNN.
of emergency medicine practitioners, particularly in determining the Two 2D image segmentation networks named U-Net (Ronneberger,
necessity of computed tomography (CT) scans. Conversely, Ortiz et al. Fischer, & Brox, 2015) and SegNet (Badrinarayanan, Kendall, & Cipolla,
(2022) developed a deep learning-driven AI model employing chest 2017) are used.
computed tomography (CT) imaging data for COVID-19 diagnosis and In addition, some filters in the structure of the original U-Net
Table 2
Some studies using DCNNs in solving medical problems.
Study Problem Dataset Description Classifier Accuracy
(Turhan, Kucuk, & Isik, Distinguish Alzheimer Disease (AD) from Alzheimer’s Disease Neuroimaging Initiative (ADNI) (2 + 1)D CNN, 3D CNN 85 % (AUC
2022) Mild Cognitive Impairment(MCI). database. (507 scans of 223 CE patients, and 507 scans score)
of 204 MCI patients
(Aljuaid, Alturki, Breast cancer classification (benign or Breast Cancer Histopathological Database (9109 breast Transfer learning of CNN 95.79 % −
Alsubaie, Cavallaro, malignant) tumor tissue images) (ResNet, InceptionV3Net, and 99.7 %
& Liotta, 2022) ShuffleNet)
(Pontoriero et al., 2021) Perform automated quality control on 200 static FDOPA PET images, 200 misalignement and Deep learning convolutional 86 % ±
(EEG and MRI) 200 low signal-to-noise ratio scan images. neural networks 0.01
(Silva-Rodríguez et al., Detection of cribriform patterns that can 182 images of prostate histology whole slide Shallow CNN 82 % (AUC
2020) help the Gleason scoring system used in score)
prostate cancer
(Chougrad, Zouaki, & Mammography mass lesion classification Digital Database of Screening Mammography (5316 Transfer learning from VGG16, 98.94 %
Alheyane, 2018) images), Breast Cancer Digital Repository (600 images), ResNet50, Inception v3
INbreast database (200 images)
(Bonavita et al., 2020) Assessment of nodule malignancy LUNA16 dataset (888 CT scans) Deep Convolutional Neural 83 %-89 %
Networks
13
network version were removed and a third network was obtained. Ex Declaration of competing interest
periments were carried out on a 3D U-Net (Çiçek, Abdulkadir, Lien
kamp, Brox, & Ronneberger, 2016), which is an improved version of U The authors declare that they have no known competing financial
Net. They were tried in a total of 4 different networks. A 99 % success interests or personal relationships that could have appeared to influence
rate has been reported with 49,094 images trained by splitting 6:2:2 as the work reported in this paper.
train:valid:test respectively.
Wang et al. (Wang, Chen, Chen, Zhong, & Zhang, 2022) proposed a Data availability
CNN model to solve the problem that the images studied on the laryn
geal disease classification problem did not have fixed resolutions. They No data was used for the research described in the article.
named their proposed model as the Hierarchical Dynamic Convolutional
Network (HDCNet), which dynamically routes the input images to small References
networks and large networks according to their resolution. Their dataset
contains 3057 images of 1950 patients. As AVG AUC performance, a Akisada, M. (1990). MIPS Standard activities in Japan. Medical Informatics, 15(1), 39–51.
https://doi.org/10.3109/14639239009025254
success of 91 % has been achieved. Aljuaid, H., Alturki, N., Alsubaie, N., Cavallaro, L., & Liotta, A. (2022). Computer-aided
diagnosis for breast cancer classification using deep neural networks and transfer
9. Conclusions learning. Comput Methods Programs Biomed, 223, Article 106951. https://doi.org/
10.1016/j.cmpb.2022.106951
Amer, M., & Maul, T. (2019). A review of modularization techniques in artificial neural
As in every field, very high-dimensional data are increasing day by networks. Artificial Intelligence Review, 52(1), 527–561. https://doi.org/10.1007/
day in the medical field. Because the data that every person in the world s10462-019-09706-7
Anderson, J. A. (1974). Diagnosis by Logistic Discriminant Function: Further Practical
creates with sensors and imaging techniques reaches enormous di Problems and Results. Applied Statistics, 23(3), 397–404. https://doi.org/10.2307/
mensions, and as the size of data increases, it becomes more difficult to 2347131
make sense of it. For this reason, the automatic extraction of valuable Appelman, F. J. R., Zuiderveld, K. J., Gash, A. G., & Viergever, M. A. (1993). Dictionary-
based medical image I/O. Medical Informatics, 18(1), 35–44. https://doi.org/
parts of the data and the interpretation of the meaning in these parts
10.3109/14639239309034466
have priority in advancing medical science and other fields of study. Arakeri, M. P., & Reddy, G. R. M. (2012). A Novel CBIR Approach to Differential
In this report, information was given about the studies that lead to Diagnosis of Liver Tumor on Computed Tomography Images. Procedia Engineering,
paradigm shifts that try to solve medical problems by using CNN models. 38, 528–536. https://doi.org/10.1016/j.proeng.2012.06.066
Ardila, D., Kiraly, A. P., Bharadwaj, S., Choi, B., Reicher, J. J., Peng, L., … Shetty, S.
Studies on CNNs structures are intensively studied in data science and (2019). End-to-end lung cancer screening with three-dimensional deep learning on
new very complex CNNs approaches are emerging. CNNs models are low-dose chest computed tomography. Nature Medicine, 25(6), 954–961. https://doi.
very complex tools and it is assumed that it will need programmer tal org/10.1038/s41591-019-0447-x
Atasever, S., Azginoglu, N., Terzi, D. S., & Terzi, R. (2023). A comprehensive survey of
ents. However, contrary to this assumption, CNNs will allow a medical deep learning research on medical image analysis with focus on transfer learning.
doctor to do their own work with much less information in the future. Clin Imaging, 94, 18–41. https://doi.org/10.1016/j.clinimag.2022.11.003
CNNs can largely disable the separate feature extraction procedures and Badrinarayanan, V., Kendall, A., & Cipolla, R. (2017). SegNet: A Deep Convolutional
Encoder-Decoder Architecture for Image Segmentation. IEEE Trans Pattern Anal Mach
handcraft engineering required for each problem. With this feature, Intell, 39(12), 2481–2495. https://doi.org/10.1109/TPAMI.2016.2644615
CNN structures will come to the fore in the near future as an artificial Batra, A. (2016). Classification of Arrhythmia Using Conjunction of Machine Learning
intelligence tool that medical doctors can use more by learning less Algorithms and ECG Diagnostic Criteria.
Batterham, P. J., Christensen, H., & Mackinnon, A. J. (2009). Modifiable risk factors
technical details. predicting major depressive disorder at four year follow-up: A decision tree
Since CNN structures offer a global solution method, we anticipate approach. BMC Psychiatry, 9(1), 75. https://doi.org/10.1186/1471-244X-9-75
that compact software will be created in the near future where people Beddiar, D. R., Oussalah, M., & Seppanen, T. (2022). Automatic captioning for medical
imaging (MIC): A rapid review of literature. Artif Intell Rev, 1–58. https://doi.org/
without engineering knowledge will be able to use these models. With
10.1007/s10462-022-10270-w
the increase in the knowledge of medical doctors on this subject, it will Beksaç, M. S., Egemen, A., Izzetoglu, K., Ergün, G., & Erkmen, A. M. (1996). An
be possible to look at the solution of new problems in computer science automated intelligent diagnostic system for the interpretation of umbilical artery
from different perspectives. Doppler velocimetry. Eur J Radiol, 23(2), 162–167. https://doi.org/10.1016/0720-
048x(96)01067-4
One of the primary challenges faced by DCNNs in the medical Binnie, C. D., Batchelor, B. G., Bowring, P. A., Darby, C. E., Herbert, L., Lloyd, D. S., …
domain is the requirement for a substantial amount of data to complete Smith, M. (1978). Computer-assisted interpretation of clinical EEGs.
their learning process, which may not always be readily available in the Electroencephalogr Clin Neurophysiol, 44(5), 575–585. https://doi.org/10.1016/0013-
4694(78)90125-6
field of medicine. To enable DCNNs to play a more active role in the Bonavita, I., Rafael-Palou, X., Ceresa, M., Piella, G., Ribas, V., & Gonzalez Ballester, M. A.
early diagnosis of rare diseases, it is essential to leverage medical (2020). Integration of convolutional neural networks for pulmonary nodule
domain expertise. This can be achieved by designing preprocessing steps malignancy assessment in a lung cancer classification pipeline. Comput Methods
Programs Biomed, 185, Article 105172. https://doi.org/10.1016/j.
that reduce the importance of patterns unrelated to the disease while cmpb.2019.105172
increasing the importance of patterns related to the disease. Addressing Burghard, C. (2021). Data and Advanced Analytics Are the New Healthcare and Life
this challenge requires medical doctors to develop a deeper under Science Currency. The Data Dilemma and Its Impact on AI in Healthcare and Life
Sciences.
standing of the operational principles of DCNNs, enabling more effective Canny, J. (1986). A computational approach to edge detection. Retrieved from IEEE
collaboration with data scientists. Trans Pattern Anal Mach Intell, 8(6), 679–698 https://www.ncbi.nlm.nih.gov/pub
Funding: Serdar Abut acknowledges a Postdoctoral grant from the med/21869365.
Chak, P., Navadiya, P., Parikh, B., & Pathak, K. C. (2019). Neural network and svm based
Scientific and Technical Research Council of Turkey (TUBITAK, 2219 -
kidney stone based medical image classification. Paper presented at the International
International Postdoctoral Research Scholarship Programme, Grant Conference on Computer Vision and Image Processing.
Number: 1059B192100853). Chao, H., Shan, H., Homayounieh, F., Singh, R., Khera, R. D., Guo, H., … Yan, P. (2021).
Ethical Approval: This article complies with ethical standards. Deep learning predicts cardiovascular disease risks from lung cancer screening low
dose computed tomography. Nature Communications, 12(1), 2963. https://doi.org/
10.1038/s41467-021-23235-4
CRediT authorship contribution statement Chodos, A., & Ouellette, J. (2001). November 8, 1895. Roentgen’s Discovery of X-rays, 10
(10), 2.
Choi, H., Yun, J. P., Lee, A., Han, S.-S., Kim, S. W., & Lee, C. (2023). Deep learning
Serdar Abut: Conceptualization, Methodology, Investigation, Re synthesis of cone-beam computed tomography from zero echo time magnetic
sources, Writing – original draft, Visualization. Hayrettin Okut: Writing resonance imaging. Scientific Reports, 13(1), 6031. https://doi.org/10.1038/s41598-
– review & editing. K. James Kallail: Supervision, Writing – review & 023-33288-8
editing.
14
Chougrad, H., Zouaki, H., & Alheyane, O. (2018). Deep Convolutional Neural Networks Kononenko, I. (2001). Machine learning for medical diagnosis: History, state of the art
for breast cancer screening. Comput Methods Programs Biomed, 157, 19–30. https:// and perspective. Artif Intell Med, 23(1), 89–109. https://doi.org/10.1016/s0933-
doi.org/10.1016/j.cmpb.2018.01.011 3657(01)00077-x
Çiçek, Ö., Abdulkadir, A., Lienkamp, S. S., Brox, T., & Ronneberger, O. (2016). 3D U-Net: Kothapalli, S. R., Yelleswarapu, C. S., Naraharisetty, S. G., Wu, P., & Rao, D. V. (2005).
Learning dense volumetric segmentation from sparse annotation. Paper presented at Spectral phase based medical image processing. Acad Radiol, 12(6), 708–721.
the International conference on medical image computing and computer-assisted. https://doi.org/10.1016/j.acra.2004.09.017
intervention. Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2017). ImageNet Classification with Deep
Dai, R. (1991). [Clinical application of image processing system and quantitative analysis Convolutional Neural Networks. Communications of the Acm, 60(6), 84–90. https://
of left ventriculogram and coronary arteriogram]. Zhonghua xin xue guan bing za zhi, doi.org/10.1145/3065386
19(2), 85-88, 124. Retrieved from http://europepmc.org/abstract/MED/1879318. Kurnaz, M. N., Dokur, Z., & Olmez, T. (2007). An incremental neural network for tissue
de Valk, J. P. J., Eijkman, E. G. J., & Vintcent, J. (1981). Detectability index, likelihood segmentation in ultrasound images. Comput Methods Programs Biomed, 85(3),
ratio and diagnostic output as psychophysical measures for evaluation of medical 187–195. https://doi.org/10.1016/j.cmpb.2006.10.010
image processing. Medical and Biological Engineering and Computing, 19(5), 597–604. Kwak, G. H., Kwak, E. J., Song, J. M., Park, H. R., Jung, Y. H., Cho, B. H., … Hwang, J. J.
https://doi.org/10.1007/BF02442774 (2020). Automatic mandibular canal detection using a deep convolutional neural
Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., & Fei-Fei, L. (2009). Imagenet: A large-scale network. Sci Rep, 10(1), 5711. https://doi.org/10.1038/s41598-020-62586-8
hierarchical image database. Paper presented at the 2009 IEEE conference on Lamba, S., Baliyan, A., & Kukreja, V. (2022). GAN based image augmentation for
computer vision and pattern recognition. increased CNN performance in Paddy leaf disease classification. Paper presented at the
Depeursinge, A., Vargas, A., Platon, A., Geissbuhler, A., Poletti, P. A., & Muller, H. 2022 2nd International Conference on Advance Computing and Innovative Technologies
(2012). Building a reference multimedia database for interstitial lung diseases. in Engineering (ICACITE).
Comput Med Imaging Graph, 36(3), 227–238. https://doi.org/10.1016/j. Lee, S., Jeong, B., Kim, M., Jang, R., Paik, W., Kang, J., … Kim, N. (2022). Emergency
compmedimag.2011.07.003 triage of brain computed tomography via anomaly detection with a deep generative
Domingues, I., Pereira, G., Martins, P., Duarte, H., Santos, J., & Abreu, P. H. (2020). model. Nature Communications, 13(1), 4251. https://doi.org/10.1038/s41467-022-
Using deep learning techniques in medical imaging: A systematic review of 31808-0
applications on CT and PET. Artificial Intelligence Review, 53(6), 4093–4160. https:// Lei, Y. M., Zhao, X. M., Wang, G. D., Yu, K. X., & Guo, W. D. (2017). A novel approach for
doi.org/10.1007/s10462-019-09788-3 cirrhosis recognition via improved LBP algorithm and dictionary learning.
Dudhane, A. A., & Talbar, S. N. (2018). Multi-scale directional mask pattern for medical Biomedical Signal Processing and Control, 38, 281–292. https://doi.org/10.1016/j.
image classification and retrieval. Paper presented at the Proceedings of 2nd bspc.2017.06.014
International Conference on Computer Vision & Image Processing. Lemke, H. U. (1993). Future directions in electronic image handling. Investigative
Eastwood, J., & Baker, T. G. (1983). Cutaneous malignant melanoma in West Yorkshire: radiology, 28(Suppl 3), S79–S81. https://doi.org/10.1097/00004424-199308003-
I. A prospective study of variables, survival and prognosis. Br J Cancer, 48(5), 00041
645–655. https://doi.org/10.1038/bjc.1983.246 Liang, C. H., Liu, Y. C., Wu, M. T., Garcia-Castro, F., Alberich-Bayarri, A., & Wu, F. Z.
Everitt, B. S., Gourlay, A. J., & Kendell, R. E. (1971). An attempt at validation of (2020). Identifying pulmonary nodules or masses on chest radiography using deep
traditional psychiatric syndromes by cluster analysis. Br J Psychiatry, 119(551), learning: External validation and strategies to improve clinical practice. Clin Radiol,
399–412. https://doi.org/10.1192/bjp.119.551.399 75(1), 38–45. https://doi.org/10.1016/j.crad.2019.08.005
Fukunaga, K. (2013). Introduction to statistical pattern recognition. Amsterdam: Elsevier. Liu, X., Du, H., Wang, G., Zhou, S., & Zhang, H. (2015). Automatic diagnosis of
Gao, S., Yang, W., Xu, M., Zhang, H., Yu, H., Qian, A., & Zhang, W. (2023). U-MLP: MLP- premature ventricular contraction based on Lyapunov exponents and LVQ neural
based ultralight refinement network for medical image segmentation. Computers in network. Comput Methods Programs Biomed, 122(1), 47–55. https://doi.org/
Biology and Medicine, 165, Article 107460. https://doi.org/10.1016/j. 10.1016/j.cmpb.2015.06.010
compbiomed.2023.107460 Lowe, D. G. (1999). Object recognition from local scale-invariant features. Paper
Gianola, D., Okut, H., Weigel, K. A., & Rosa, G. J. (2011). Predicting complex presented at the Proceedings of the seventh IEEE international conference on computer
quantitative traits with Bayesian neural networks: A case study with Jersey cows and vision.
wheat. BMC Genet, 12(1), 87. https://doi.org/10.1186/1471-2156-12-87 Ma, P., Li, C., Rahaman, M. M., Yao, Y., Zhang, J., Zou, S., … Grzegorzek, M. (2022).
Gonzales, R. C. (1987). Digital Image Processing (2nd ed.). New York: Addison Wesley. A state-of-the-art survey of object detection techniques in microorganism image
Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep learning. Massachusetts: MIT analysis: From classical methods to deep learning approaches. Artif Intell Rev, 1–72.
press. https://doi.org/10.1007/s10462-022-10209-1
Grossi, E., Olivieri, C., & Buscema, M. (2017). Diagnosis of autism through EEG processed Maroulis, D. E., Iakovidis, D. K., Karkanis, S. A., & Karras, D. A. (2003). CoLD: A versatile
by advanced computational algorithms: A pilot study. Comput Methods Programs detection system for colorectal lesions in endoscopy video-frames. Comput Methods
Biomed, 142, 73–79. https://doi.org/10.1016/j.cmpb.2017.02.002 Programs Biomed, 70(2), 151–166. https://doi.org/10.1016/s0169-2607(02)00007-x
Heckel, F., Meine, H., Moltz, J. H., Kuhnigk, J. M., Heverhagen, J. T., Kiessling, A., … Matsuzaka, Y., & Yashiro, R. (2022). Applications of Deep Learning for Drug Discovery
Hahn, H. K. (2014). Segmentation-based partial volume correction for volume Systems with BigData. BioMedInformatics, 2(4), 603–624. https://doi.org/10.3390/
estimation of solid lesions in CT. IEEE Trans Med Imaging, 33(2), 462–480. https:// biomedinformatics2040039
doi.org/10.1109/tmi.2013.2287374 Michael, D. J., & Nelson, A. C. (1989). HANDX: A model-based system for automatic
Heo, S., Ha, J., Jung, W., Yoo, S., Song, Y., Kim, T., & Cha, W. C. (2022). Decision effect segmentation of bones from digital hand radiographs. IEEE Transactions on Medical
of a deep-learning model to assist a head computed tomography order for pediatric Imaging, 8(1), 64–69. https://doi.org/10.1109/42.20363
traumatic brain injury. Scientific Reports, 12(1), 12454. https://doi.org/10.1038/ Michel, J.-B., Shen, Y. K., Aiden, A. P., Veres, A., Gray, M. K., Pickett, J. P., … Aiden, E. L.
s41598-022-16313-0 (2011). Quantitative Analysis of Culture Using Millions of Digitized Books., 331(6014),
Hsu, K., Yuh, D.-Y., Lin, S.-C., Lyu, P.-S., Pan, G.-X., Zhuang, Y.-C., … Juan, C.-J. (2022). 176–182. https://doi.org/10.1126/science.1199644
Improving performance of deep learning models using 3.5D U-Net via majority Mitchell, T. M. (1997). Machine Learning (Vol. 1). New York: McGraw-Hill.
voting for tooth segmentation on cone beam computed tomography. Scientific Monroy, B., Sanchez, K., Arguello, P., Estupiñán, J., Bacca, J., Correa, C. V., … Rojas-
Reports, 12(1), 19809. https://doi.org/10.1038/s41598-022-23901-7 Morales, F. (2023). Automated chronic wounds medical assessment and tracking
Hu, Y. L., Rogers, W. J., Coast, D. A., Kramer, C. M., & Reichek, N. (1998). Vessel framework based on deep learning. Computers in Biology and Medicine, 165, Article
boundary extraction based on a global and local deformable physical model with 107335. https://doi.org/10.1016/j.compbiomed.2023.107335
variable stiffness. Magn Reson Imaging, 16(8), 943–951. https://doi.org/10.1016/ Moody, G. B., & Mark, R. G. (2001). The impact of the MIT-BIH arrhythmia database.
s0730-725x(98)00105-2 IEEE Eng Med Biol Mag, 20(3), 45–50. https://doi.org/10.1109/51.932724
Huda, W., & Abrahams, R. B. (2015). X-ray-based medical imaging and resolution. AJR Murtaza, G., Shuib, L., Wahab, A. W. A., Mujtaba, G., Nweke, H. F., Al-garadi, M. A., …
Am J Roentgenol, 204(4), W393–W397. https://doi.org/10.2214/AJR.14.13126 Azmi, N. A. (2020). Deep learning-based breast cancer classification through medical
Ichikawa, S., Hamada, M., & Sugimori, H. (2021). A deep-learning method using imaging modalities: State of the art and research challenges. Artificial Intelligence
computed tomography scout images for estimating patient body weight. Scientific Review, 53(3), 1655–1720. https://doi.org/10.1007/s10462-019-09716-5
Reports, 11(1), 15627. https://doi.org/10.1038/s41598-021-95170-9 Nishio, M., & Nagashima, C. (2017). Computer-aided Diagnosis for Lung Cancer:
Inglis, R., Windolf, J., & Pannike, A. (1990). Medical image processing using a personal Usefulness of Nodule Heterogeneity. Acad Radiol, 24(3), 328–336. https://doi.org/
computer; digital bone density measurement with a new non-invasive procedure. 10.1016/j.acra.2016.11.007
Biomed Tech (Berl), 35(Suppl 2), 186–187. Nomura, Y., Hoshiyama, M., Akita, S., Naganishi, H., Zenbutsu, S., Matsuoka, A., …
Islam, K. T., Wijewickrema, S., & O’Leary, S. (2022). A Deep Learning Framework for Mitsukawa, N. (2023). Computer-aided diagnosis for screening of lower extremity
Segmenting Brain Tumors Using MRI and Synthetically Generated CT Images. lymphedema in pelvic computed tomography images using deep learning. Scientific
Sensors (Basel), 22(2). https://doi.org/10.3390/s22020523 Reports, 13(1), 16214. https://doi.org/10.1038/s41598-023-43503-1
Jaskari, J., Sahlsten, J., Järnstedt, J., Mehtonen, H., Karhu, K., Sundqvist, O., … Kaski, K. Noz, M. E., Maguire, G. Q., Birnbaum, B. A., Sanger, J. J., Kramer, E. L., Chapnick, J. V.,
(2020). Deep Learning Method for Mandibular Canal Segmentation in Dental Cone & Kaminer, E. A. (1993). Graphical interface for medical image processing. Journal of
Beam Computed Tomography Volumes. Scientific Reports, 10(1), 5842. https://doi. Medical Systems, 17(1), 1–16. https://doi.org/10.1007/BF01000582
org/10.1038/s41598-020-62321-3 Odaka, T., Takahama, T., Wagatsuma, H., Shimada, K., & Ogura, H. (1994). A visual data
Jensen, P. B., Jensen, L. J., & Brunak, S. (2012). Mining electronic health records: analysis system for the medical image processing. Journal of Medical Systems, 18(3),
Towards better research applications and clinical care. Nat Rev Genet, 13(6), 151–157. https://doi.org/10.1007/BF00997293
395–405. https://doi.org/10.1038/nrg3208 Okut, H. (2016). Bayesian Regularized Neural Networks for Small n Big p Data. In G. R.
Kim, J. H., Kim, H. Y., & Lee, Y. S. (2001). A novel method using edge detection for signal Joao Luis (Ed.), Artificial Neural Networks - Models and Applications (pp. Ch. 2).
extraction from cDNA microarray image analysis. Exp Mol Med, 33(2), 83–88. Rijeka: IntechOpen.
https://doi.org/10.1038/emm.2001.15
15
Okut, H. (2021). Deep Learning Algorithms for Complex Traits Genomic Prediction. pulmonary angiogram: A systematic review and meta-analysis. Scientific Reports, 11
Journal of Animal Science and Products (JASP), 4(2), 225–239. https://doi.org/ (1), 15814. https://doi.org/10.1038/s41598-021-95249-3
10.51970/jasp.1039713 Soomro, T. A., Zheng, L., Afifi, A. J., Ali, A., Yin, M., & Gao, J. (2022). Artificial
Okut, H., Gianola, D., Rosa, G. J., & Weigel, K. A. (2011). Prediction of body mass index intelligence (AI) for medical imaging to combat coronavirus disease (COVID-19): A
in mice using dense molecular markers and a regularized neural network. Genet Res detailed review with direction for future research. Artif Intell Rev, 55(2), 1409–1439.
(Camb), 93(3), 189–201. https://doi.org/10.1017/S0016672310000662 https://doi.org/10.1007/s10462-021-09985-z
Okut, H., Wu, X. L., Rosa, G. J., Bauck, S., Woodward, B. W., Schnabel, R. D., … Stachowiak, G. W., Wolski, M., Woloszynski, T., Podsiadlo, P. J. B., & Biotribology..
Gianola, D. (2013). Predicting expected progeny difference for marbling score in (2016). Detection and prediction of osteoarthritis in knee and hand joints based on the X-
Angus cattle using artificial neural networks and Bayesian regression models. Genet ray image analysis., 2(4), 162–172.
Sel Evol, 45(1), 34. https://doi.org/10.1186/1297-9686-45-34 Sun, H., Wang, F., Yang, Y., Hong, X., Xu, W., Wang, S., … Lu, L. (2023). Transfer
Ortiz, A., Trivedi, A., Desbiens, J., Blazes, M., Robinson, C., Gupta, S., … Ferres, J. M. L. learning-based attenuation correction for static and dynamic cardiac PET using a
(2022). Effective deep learning approaches for predicting COVID-19 outcomes from generative adversarial network. European Journal of Nuclear Medicine and Molecular
chest computed tomography volumes. Scientific Reports, 12(1), 1716. https://doi. Imaging, 50(12), 3630–3646. https://doi.org/10.1007/s00259-023-06343-9
org/10.1038/s41598-022-05532-0 Sutou, Y., & Kahara, S. (1975). Method of feature extraction in image processing and its
Palkovics, D., Mangano, F. G., Nagy, K., & Windisch, P. (2020). Digital three-dimensional application for medical image processing. Iyodenshi To Seitai Kogaku, 13(1), 9–17.
visualization of intrabony periodontal defects for regenerative surgical treatment Turhan, G., Kucuk, H., & Isik, E. O. (2022). Spatio-temporal convolution for classification
planning. BMC Oral Health, 20(1), 351. https://doi.org/10.1186/s12903-020- of alzheimer disease and mild cognitive impairment. Comput Methods Programs
01342-w Biomed, 221, Article 106825. https://doi.org/10.1016/j.cmpb.2022.106825
Pastorino, R., De Vito, C., Migliara, G., Glocker, K., Binenbaum, I., Ricciardi, W., & Udupa, J. K., & Goncalves, R. J. (1993). Medical image rendering. Retrieved from
Boccia, S. (2019). Benefits and challenges of Big Data in healthcare: An overview of American journal of cardiac imaging, 7(3), 154–163 http://europepmc.org/abstract
the European initiatives. Eur J Public Health, 29, 23–27. https://doi.org/10.1093/ /MED/10146371.
eurpub/ckz168 Udupa, J. K., Srihari, S. N., & Herman, G. T. (1982). Boundary Detection in
Pauker, S. G., & Kassirer, J. P. (1975). Therapeutic decision making: A cost-benefit Multidimensions. IEEE Transactions on Pattern Analysis and Machine Intelligence,
analysis. N Engl J Med, 293(5), 229–234. https://doi.org/10.1056/ PAMI-4(1), 41–50. https://doi.org/10.1109/TPAMI.1982.4767193
NEJM197507312930505 Van Der Maaten, L., Postma, E., & Herik, J.V.. (2009). Dimensionality reduction: a.
Pawlak, Z. (1982). Rough sets. International Journal of Computer & Information Sciences, comparative review., 10(66–71), 13.
11(5), 341–356. https://doi.org/10.1007/BF01001956 Vinayahalingam, S., Xi, T., Berge, S., Maal, T., & de Jong, G. (2019). Automated
Pontoriero, A. D., Nordio, G., Easmin, R., Giacomel, A., Santangelo, B., Jahuar, S., … detection of third molars and mandibular nerve by deep learning. Sci Rep, 9(1),
Veronese, M. (2021). Automated Data Quality Control in FDOPA brain PET Imaging 9007. https://doi.org/10.1038/s41598-019-45487-3
using Deep Learning. Comput Methods Programs Biomed, 208, Article 106239. Wang, S., Chen, Y., Chen, S., Zhong, Q., & Zhang, K. (2022). Hierarchical dynamic
https://doi.org/10.1016/j.cmpb.2021.106239 convolutional neural network for laryngeal disease classification. Sci Rep, 12(1),
Raj, J. R. F., Vijayalakshmi, K., & Priya, S. K. (2019). Medical image denoising using 13914. https://doi.org/10.1038/s41598-022-18217-5
multi-resolution transforms. Measurement, 145, 769–778. https://doi.org/10.1016/j. Wiegand, D. A., Page, R. B., & Channin, D. S. (1993). The surgical workstation: Surgical
measurement.2019.01.001 planning using generic software. Otolaryngology-Head and Neck Surgery, 109(3),
Rajesh, T., & Malar, R. S. M. (2013). Rough set theory and feed forward neural network 434–440. https://doi.org/10.1177/019459989310900308
based brain tumor detection in magnetic resonance images. Paper presented at the Williams, D. M., Bland, P., Liu, L., Farjo, L., Francis, I. R., & Meyer, C. R. (1989). Liver-
International Conference on Advanced Nanomaterials & Emerging Engineering. tumor boundary detection: Human observer vs computer edge detection. Invest
Technologies. Radiol, 24(10), 768–775. https://doi.org/10.1097/00004424-198910000-00008
Rompianesi, G., Pegoraro, F., Ceresa, C. D., Montalti, R., & Troisi, R. I. (2022). Artificial Wu, B., Li, C., Zhang, J., Lai, H., Feng, Q., & Huang, M. (2023). Unsupervised dual-
intelligence in the diagnosis and management of colorectal cancer liver metastases. domain disentangled network for removal of rigid motion artifacts in MRI. Computers
World J Gastroenterol, 28(1), 108–122. https://doi.org/10.3748/wjg.v28.i1.108 in Biology and Medicine, 165, Article 107373. https://doi.org/10.1016/j.
Ronneberger, O., Fischer, P., & Brox, T. (2015). U-net: Convolutional networks for compbiomed.2023.107373
biomedical image segmentation. Paper presented at the International Conference on Wu, P.-Y., Cheng, C.-W., Kaddi, C. D., Venugopalan, J., Hoffman, R., & Wang, M. D.
Medical image computing and computer-assisted intervention. (2016). Omic and electronic health record big data analytics for precision medicine.
Safdarian, N., & Dabanloo, N. J. (2021). Detection and Classification of COVID-19 by IEEE Transactions on Biomedical Engineering, 64(2), 263–273.
Lungs Computed Tomography Scan Image Processing using Intelligence Algorithm. Wu, P. Y., Cheng, C. W., Kaddi, C. D., Venugopalan, J., Hoffman, R., & Wang, M. D.
J Med Signals Sens, 11(4), 274–284. https://doi.org/10.4103/jmss.JMSS_55_20 (2017). Omic and Electronic Health Record Big Data Analytics for Precision
Saito, M., & Hayashi, T. (1982). An approach to inverse scattering problems. Medicine. IEEE Trans Biomed Eng, 64(2), 263–273. https://doi.org/10.1109/
Bioelectromagnetics, 3(1), 73–80. https://doi.org/10.1002/bem.2250030111 tbme.2016.2573285
Samei, E. (2005). AAPM/RSNA physics tutorial for residents: Technological and Xu, H., Chen, X., Qian, P., & Li, F. (2023). A two-stage segmentation of sublingual veins
psychophysical considerations for digital mammographic displays. Radiographics, 25 based on compact fully convolutional networks for Traditional Chinese Medicine
(2), 491–501. https://doi.org/10.1148/rg.252045185 images. Health Information Science and Systems, 11(1), 19. https://doi.org/10.1007/
Schwartz, F. R., Clark, D. P., Rigiroli, F., Kalisz, K., Wildman-Tobriner, B., Thomas, S., … s13755-023-00214-1
Marin, D. (2023). Evaluation of the impact of a novel denoising algorithm on image Xu, J., Ren, H., Cai, S., & Zhang, X. (2023). An improved faster R-CNN algorithm for
quality in dual-energy abdominal CT of obese patients. European Radiology, 33(10), assisted detection of lung nodules. Comput Biol Med, 153, Article 106470. https://
7056–7065. https://doi.org/10.1007/s00330-023-09644-7 doi.org/10.1016/j.compbiomed.2022.106470
Semmlow, J., Hand, W., Ackerman, L. V., & Alcorn, F. (1980). Computer-controlled Yang, H. X., Shan, C. F., Kolen, A. F., & de With, P. H. N. (2022). Medical instrument
access device for medical image processing. Retrieved from Medical instrumentation, detection in ultrasound: A review. Artificial Intelligence Review, 1–40. https://doi.org/
14(3), 184–186 http://europepmc.org/abstract/MED/7382904. 10.1007/s10462-022-10287-1
Shailaja, K., Seetharamulu, B., & Jabbar, M. A. (2018). Machine Learning in Healthcare: Yang, X., Park, S., Lee, S., Han, K., Lee, M. R., Song, J. S., … Do Yang, J. (2023).
A Review. Paper presented at the 2018 Second International Conference on Electronics, Estimation of right lobe graft weight for living donor liver transplantation using deep
Communication and Aerospace Technology (ICECA). learning-based fully automatic computed tomographic volumetry. Scientific Reports,
Shi, Z., Miao, C., Schoepf, U. J., Savage, R. H., Dargis, D. M., Pan, C., … Zhang, L. J. 13(1), 17746. https://doi.org/10.1038/s41598-023-45140-0
(2020). A clinically applicable deep-learning model for detecting intracranial Yuan, J., Zhou, F., Guo, Z., Li, X., & Yu, H. (2023). HCformer: Hybrid CNN-Transformer
aneurysm in computed tomography angiography images. Nature Communications, 11 for LDCT Image Denoising. Journal of Digital Imaging, 36(5), 2290–2305. https://doi.
(1), 6090. https://doi.org/10.1038/s41467-020-19527-w org/10.1007/s10278-023-00842-9
Silva-Rodríguez, J., Colomer, A., Sales, M. A., Molina, R., Naranjo, V. J. C. M., Zeleznik, R., Foldyna, B., Eslami, P., Weiss, J., Alexander, I., Taron, J., …
Biomedicine, P., & i.. (2020). Going deeper through the Gleason scoring scale: An Aerts, H. J. W. L. (2021). Deep convolutional neural networks to predict
automatic end-to-end system for histology prostate grading and cribriform pattern cardiovascular risk from computed tomography. Nature Communications, 12(1), 715.
detection., 195, Article 105637. https://doi.org/10.1038/s41467-021-20966-2
Sinzinger, F., Astaraki, M., Smedby, Ö., & Moreno, R. (2022). Spherical Convolutional Zhang, F., Wang, L., Zhao, J., & Zhang, X. (2023). Medical applications of generative
Neural Networks for Survival Rate Prediction in Cancer Patients. Front Oncol, 12, adversarial network: A visualization analysis. Acta Radiologica, 64(10), 2757–2767.
Article 870457. https://doi.org/10.3389/fonc.2022.870457 https://doi.org/10.1177/02841851231189035
Sobel, I., & Feldman, G. (1973). A 3×3 isotropic gradient operator for image processing. Zhang, J., Li, C., Rahaman, M. M., Yao, Y., Ma, P., Zhang, J., … Grzegorzek, M. (2022).
Pattern Classification and Scene Analysis, 271–272. A comprehensive review of image analysis methods for microorganism counting:
Soffer, S., Klang, E., Shimon, O., Barash, Y., Cahan, N., Greenspana, H., & Konen, E. From classical image processing to deep learning approaches. Artif Intell Rev, 55(4),
(2021). Deep learning for pulmonary embolism detection on computed tomography 2875–2944. https://doi.org/10.1007/s10462-021-10082-4
16

RPL 211

Uploaded by

Copyright:

Available Formats

RPL 211

Uploaded by

Document Information

Original Title

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

RPL 211

Uploaded by

Copyright:

Available Formats

Expert Systems With Applications 244 (2024) 122983

Contents lists available at ScienceDirect

Expert Systems With Applications

Paradigm shift from Artificial Neural Networks (ANNs) to deep

1. Introduction accumulation voluminous complex healthcare data contain ample in

Fig. 2. Electronic health recods and their analysis procedure.

ADMET Absorption, Distribution, Metabolism And ExcretionToxicology.

Fig. 3. A Hierarchical Representation of Some Artificial Intelligence Concepts.

(Dai, 1991) focused on clinical applications in China using the “IA-87

Fig. 7. Convolution of 3x3 kernel matrix to a 10x10 image matrix.

demographic and clinical background, the genetic architecture of dis

6. A paradigm shift from ANNs to DCNNs

Fig. 11 is provided to visually demonstrate the differentiation in how

You might also like

RPL 211

Uploaded by

Copyright:

Available Formats

RPL 211

Uploaded by

Document Information

Original Title

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

RPL 211

Uploaded by

Copyright:

Available Formats

Expert Systems With Applications 244 (2024) 122983

Contents lists available at ScienceDirect

Expert Systems With Applications

Paradigm shift from Artificial Neural Networks (ANNs) to deep

1. Introduction accumulation voluminous complex healthcare data contain ample in­

Fig. 2. Electronic health recods and their analysis procedure.

ADMET Absorption, Distribution, Metabolism And ExcretionToxicology.

Fig. 3. A Hierarchical Representation of Some Artificial Intelligence Concepts.

(Dai, 1991) focused on clinical applications in China using the “IA-87

Fig. 7. Convolution of 3x3 kernel matrix to a 10x10 image matrix.

demographic and clinical background, the genetic architecture of dis­

6. A paradigm shift from ANNs to DCNNs

Fig. 11 is provided to visually demonstrate the differentiation in how

You might also like

1. Introduction accumulation voluminous complex healthcare data contain ample in

demographic and clinical background, the genetic architecture of dis