Healthcare Analytics 4 (2023) 100216

Contents lists available at ScienceDirect

Healthcare Analytics
journal homepage: www.elsevier.com/locate/health

A comprehensive review of deep neural networks for medical image

processing: Recent developments and future opportunities
Pawan Kumar Mall a ,∗, Pradeep Kumar Singh b , Swapnita Srivastav c , Vipul Narayan c ,
Marcin Paprzycki d , Tatiana Jaworska d , Maria Ganzha d
a
Lovely Professional University, Jalandhar - Delhi G.T. Road, Phagwara, Punjab 144411, India
b
Department of Computer Science and Engineering, Madan Mohan Malaviya University of Technology, Gorakhpur, U.P, 273008, India
c Galgotias University, Greater Noida, India
d Systems Research Institute, Polish Academy of Sciences, Poland

ARTICLE INFO ABSTRACT

Keywords: Artificial Intelligence (AI) solutions have been widely used in healthcare, and recent developments in deep
Artificial intelligence neural networks have contributed to significant advances in medical image processing. Much ongoing research
Deep neural networks is aimed at helping medical practitioners by providing automated systems to analyze images and diagnose acute
Predictive analytics
diseases, such as brain tumors, bone cancer, breast cancer, bone fracture, and many others. This comprehensive
Machine learning
review delivers an overview of recent advances in medical imaging using deep neural networks. In addition to
Medical imaging diagnostic analytics
the comprehensive literature review, a summary of openly available data sources and future research directions
are outlined.

1. Introduction Deep neural networks (DNNs) have significantly improved diagno-

sis, treatment planning, and patient care by revolutionizing several
Recent years can be characterized by rapid developments in the area facets of medical image processing. Their capacity to extract signifi-
of, broadly understood, artificial intelligence (AI). One of the subareas cant characteristics and patterns from medical image using large-scale
of AI that has reached real-world applicability is machine learning datasets has demonstrated to be very successful, resulting in more
(ML). Here, some of the key developments involve neural networks precise and effective analysis [3]. DNNs are especially effective at
(NN). It can be said that the development of artificial neural networks tasks like image classification and segmentation, which are essential in
proceeds in a sinusoidal fashion. After an early interest in the late medical imaging. Deep networks may learn to recognize and categorize
1950s and early 1960s, a period of relative dormancy lasted till 1986, different anatomical features, lesions, or anomalies in medical image by
when James McClelland and David Rumelhart published their famous training on massive annotated datasets. Additionally, they can precisely
book [1]. This revived interest in neural network research. However, divide up organs or other areas of interest, allowing for exact mea-
by the end of the 20th century, the interest in a neural networks has surements and quantitative analysis. DNNs have been heavily included
diminished again. One of the reasons for the stagnation was a lack of into computer-aided diagnostic (CAD) systems [4]. These networks may
computer hardware capable of handling the amount of data needed identify minor patterns or irregularities in medical images that may
to make obtainable neural network models actually applicable [2]. It be difficult for human observers to notice by learning from enormous
is only in the last ten years that renewed interest in NNs resulted volumes of labeled data. DNN-based CAD solutions benefit radiolo-
in the development of successful applications capable of addressing gists and doctors by supplying more details and enhancing diagnostic
actual, real-life problems. Some of the neural network architectures that precision. Deep learning algorithms have been effectively used for
attracted a lot of interest in deep neural networks (DNNs). They have image restoration and enhancement jobs in the field of medicine [5].
been implemented in many areas, such as medical image classification, For instance, DNNs may recreate high-quality image from imperfect
electromyography recognition, disease recognition and segmentation, or noisy data in computed tomography (CT) and magnetic resonance
and many others. However, in this work, we are interested in providing imaging (MRI), lowering artifacts and enhancing image quality. These
an overview of the use of DNNs in medical imaging. developments result in quicker scans, lower radiation doses, and better

∗ Corresponding author.
E-mail addresses: [email protected] (P.K. Mall), [email protected] (P.K. Singh), [email protected] (S. Srivastav),
[email protected] (V. Narayan), [email protected] (M. Paprzycki), [email protected] (T. Jaworska),
[email protected] (M. Ganzha).

https://doi.org/10.1016/j.health.2023.100216
Received 29 April 2023; Received in revised form 16 June 2023; Accepted 21 June 2023

2772-4425/© 2023 The Author(s). Published by Elsevier Inc. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
P.K. Mall, P.K. Singh, S. Srivastav et al. Healthcare Analytics 4 (2023) 100216

anatomical structure visualization. Deep neural networks have been tomography (CT), positron emission tomography (PET), ultrasound,
employed for accurate and robust image registration, which involves and X-ray, among others. Each modality provides unique informa-
aligning multiple medical images acquired from different modalities tion about the anatomical structures or physiological processes being
or time points [6]. By learning spatial transformations, DNNs can studied.
automatically align images and facilitate comparisons, such as tracking The DNN models are trained using large datasets of labeled medical
disease progression or planning interventions. Disease Detection and images corresponding to the specific modality or modalities being used.
Prediction: DNNs have shown promise in automated disease detection These datasets are crucial for the DNNs to learn patterns and features
and prediction using medical images. By learning from large-scale relevant to the desired tasks. The training process involves optimizing
datasets, these networks can identify specific imaging biomarkers as- the network’s parameters through techniques like backpropagation
sociated with diseases. For example, in cancer imaging, DNNs can and gradient descent, allowing the model to learn how to accurately
identify tumor characteristics, predict tumor malignancy, or assess classify, segment, or detect specific features in the medical images [11].
treatment response based on radiological images. Deep neural networks Once the DNN models are trained, they can be used for various
can generate synthetic medical images, aiding in data augmentation, analyses on unseen medical images. This can include tasks such as
training set expansion, and creating realistic simulations for training image classification, where the DNN identifies the presence or absence
and testing purposes. This synthesis of images is particularly useful in of certain anatomical or pathological features. It can also involve image
situations with limited annotated data or rare conditions, where DNNs segmentation, where the DNN identifies and delineates specific struc-
can generate diverse examples to enhance the performance of models. tures or regions within the image. Additionally, DNNs can be employed
Personalized Medicine and deep learning models can analyze medi- for object detection, where they locate and localize multiple instances
cal images in conjunction with other clinical data, such as genomics of specific objects or abnormalities within the image.
or electronic health records, to enable personalized medicine. By in- The detailed study [12] on GAN-based medical image enhancement.
tegrating patient-specific information, DNNs can assist in treatment The problem of small training sample sizes for medical image diagnostic
planning, prognostication, and selection of optimal therapies based on and treatment models is well solved by this strategy. The amount
image-derived features and patterns. of training samples for models of medical image-based diagnosis and
The main structure of this article is: Section 2 an overview of therapy is growing as deep learning progresses. Due to their exceptional
deep neural network. Section 3 gives details about Medical image picture-generating capabilities and extensive usage in data augmen-
datasets organized according to organs. Section 4 gives details Medical tation, Generative Adversarial Networks (GANs) have drawn interest
image datasets organized according to organs. Section 5 states the in medical image processing. In [13], authors have focuses on three
researchChallenges and Future Directions Section 6 states the research healthcare imaging modalities and organs, including cardiac cine-MRI,
conclusion of this article. liver CT, and RGB retina pictures, we investigated several GAN designs,
ranging from straightforward DCGAN to more complex style-based
2. Deep neural networks GANs. To assess the visual acuity of the pictures produced by GANs,
their FID scores were computed using training data from well-known
In general, a deep neural network (DNN) is a subset of Machine and frequently used datasets. By assessing the segmentation accuracy
Learning (ML) that uses large multilayer neural networks. The key ideas of a U-Net trained on these produced pictures and the original data, we
behind these networks materialize in the form of algorithms, inspired further evaluated their use. The findings show that GANs are not cre-
by the way that the human brain deals with large amounts of data ated equal, with some doing poorly in applications involving medical
(for training). DNN algorithms perform a task over and over again, imaging while others fared substantially better. The best GANs are able
eventually improving the results as a result of the deep layers that to produce medical pictures that are convincing by FID standards, pass
allow progressive learning to occur. DNN is an architecture with a certain criteria, and deceive skilled professionals in a visual Turing test.
large number of, usually wide, layers between the input and output The findings of segmentation, however, imply that no GAN is able to
layers. For instance, currently, DNNs can have beyond 1200 layers replicate the complete richness of medical information.
and more than 16 million neurons [7]. Moreover, while these layers Deep neural networks are now a beneficial technology for medical
are not homogeneous, they all consist of neurons whose function is imaging workflow optimization and clinical care enhancement. So far,
modeled after actual neurons by an activation function (non-linear it has triggered a massive global interest and has gradually delivered
or linear) [8], such as hyperbolic tangent, sigmoid, etc. There are promising outcomes when applied to multiple fields [14,15]. Deep
some basic layers according to the roles that they have in different learning’s development in medical image analysis follows a slower
architectures. The layer where the extraction of features takes place is but comparable path to computer vision. However, using computer
a convolutional layer, the name of which comes from the convolution vision techniques directly may not produce adequate results due to the
operations performed using different filters. There exists a large number differences between medical pictures and natural images. It is necessary
of specific DNN architectures [9]. Let us summarize the main ones. to solve difficulties specific to medical imaging jobs in order to obtain
The interested reader should consult the cited works for more detail, optimal performance [16].
including a representation of each architecture. The hierarchical representation in Fig. 1 likely demonstrates how
Deep neural networks (DNNs) have indeed been widely used to these trained DNN models can be applied to different stages of medical
improve the efficiency of pattern recognition tasks in various organs image analysis, starting from the raw input images and progressing
or anatomical regions within medical imaging. These DNN models towards more detailed and specific tasks. This hierarchical approach
have shown promising results in areas such as kidney, liver, brain, allows for the extraction of meaningful information from medical im-
bone, breast, chest, eyes, and more. They have the potential to provide ages and facilitates accurate and efficient pattern recognition.Let us
accurate and automated analysis of medical images, aiding in diagnosis, now summarize the main DNN architectures reported in the literature
treatment planning, and monitoring of diseases [10]. in the context of medical image processing
The hierarchical representation of medical images and analyses
using deep neural networks typically involves several stages. Fig. 1 2.1. Classification based on Convolutional Neural Networks (ConvNet/
(not provided) likely illustrates this hierarchy, which I can describe in CNN)
general terms.
At the bottom row of the hierarchy, different types of modalities are This network structure was first described in [17] in 1989. Two
used for training the DNN models. These modalities can include imag- essential parts of the CNN architecture are feature extractors and clas-
ing techniques such as magnetic resonance imaging (MRI), computed sification [18]. Convolution operations with different masks extract a

2
P.K. Mall, P.K. Singh, S. Srivastav et al. Healthcare Analytics 4 (2023) 100216

Fig. 1. Hierarchical representation of medical image analysis, in which neural networks have been applied.

set of features based on which classification is conducted. The ConvNet The significance of CNNs in disease detection lies in their ability
consists of three elementary types of layers: convolution, max-pooling, to leverage the power of deep learning, extracting meaningful and
fully connected (FC), and softmax for classification. discriminative features from medical images. By improving detection
Several essential processes are employed in convolutional neural accuracy, reducing interpretation time, and providing consistent re-
networks (CNNs) for image classification, including convolution, max- sults, CNNs contribute to more effective and efficient disease diagnosis,
pooling, fully connected (FC) layers, and softmax activation. Here enabling early intervention, better patient outcomes, and enhanced
is a description of each of these elements within the framework of healthcare decision-making. This paper [20] gives an overview of
CNN-based classification: the literature in the area of deep convolutional neural network-based
Convolution: The primary feature extraction algorithm used by medical picture segmentation. The article looks at a number of popular
CNNs is convolution. It entails applying to the input picture a collection medical picture datasets, several segmentation task evaluation mea-
of teachable filters, often known as kernels. Between its weights and the sures, and the effectiveness of various CNN-based networks. In [21]
input’s local receptive field, each filter performs a dot product. Edges, this article, authors have suggested a unique and computationally
corners, and other little patterns or characteristics inside the image can efficient deep learning strategy to exploit minimal data for learning
be found with this technique. A series of feature maps, each of which domain invariant and generalizable representations in several medical
reflects the response of a particular filter over the whole input picture, imaging applications, such as malaria, diabetic retinopathy, and TB.
are the result of the convolutional layer. The CNNs produced by our method are referred to as Incremental
Max-Pooling: Max-pooling is used to minimize the spatial dimen- Modular Networks (IMNets), and we call our methodology Incremental
sions and downsample the feature maps. Within each feature map, it Modular Network Synthesis (IMNS). Small network modules that we re-
acts on tiny rectangular sections (usually 2 × 2 or 3 × 3) and produces fer to as SubNets, which are capable of producing salient characteristics
the greatest value therein. Max-pooling aids in preserving the most for a specific challenge, are the foundation of our IMNS strategy. Then,
important aspects while removing unnecessary or duplicate information
by merging these SubNets in various configurations, we create networks
by choosing the maximum activation. Additionally, it adds a certain
that are increasingly larger and more potent. Only one new SubNet
amount of translational invariance, strengthening the network’s resis-
module is subject to learning updates at each level. This facilitates
tance to minute spatial fluctuations. Despite having smaller spatial
network optimization and lowers the computing resource needs for
dimensions, the resulting pooled feature maps nevertheless include
training.
vital details.
Softmax Activation: A softmax layer is often the last layer of the
2.2. DNNs in segmentation based on U-NN and V-NN architecture
CNN and is utilized for classification. The previous FC layer’s outputs
are transformed into probabilities by Softmax. When the softmax func-
tion is used, the values are normalized and converted into a probability Deep neural networks (DNNs) have significantly advanced the field
distribution. The softmax function makes sure that the probabilities add of medical image segmentation. Segmentation plays a crucial role in
up to 1, and each output neuron belongs to a certain class. By choosing extracting precise information from medical images, such as identi-
the class with the highest probability, the predicted class is established. fying and delineating specific anatomical structures or pathological
Convolutional layers are used by CNNs to extract local features, regions [22]. Here’s how DNNs contribute to medical image segmen-
max-pooling layers to downsample and preserve important informa- tation:
tion, fully connected layers to capture global relationships, and a Feature Learning: DNNs are capable of automatically learning com-
softmax activation layer for classification by supplying probability dis- plex and discriminative features directly from the input images. This
tributions over various classes. CNNs can successfully learn hierarchical eliminates the need for manual feature engineering and allows the
representations from pictures and produce precise predictions because network to adapt to different imaging modalities and anatomical vari-
to this mix of activities. ations [23]. Deep architectures, such as convolutional neural networks
Fig. 2 demonstrates the basic ConvNet architecture for the classi- (CNNs), can capture both low-level features (e.g., edges, textures) and
fication problem, which has been modified in terms of class numbers high-level semantic information (e.g., shapes, structures) through mul-
or/and layer types. Researchers further strengthened CNNs by increas- tiple layers of non-linear transformations. This feature learning ability
ing additional numbers of convolutional layers and other, for instance, enables DNNs to better represent the characteristics of medical images
dropout layers and published different state-of-the-art research work on and improve segmentation accuracy [24].
various computer vision problems. For medical application details, see End-to-End Learning: DNNs enable end-to-end learning, where the
Section 4. entire segmentation process is learned from the raw input images.

3
P.K. Mall, P.K. Singh, S. Srivastav et al. Healthcare Analytics 4 (2023) 100216

Fig. 2. CNN architecture for classification.

Fig. 3. U-Net DNN framework by [19]. Generally, the left half of this structure is called a decoder, and the right half of this structure is called an encoder.

Traditional segmentation methods often involve multi-step pipelines, DNNs. These techniques may involve applying morphological opera-
including preprocessing, feature extraction, and classification [25]. tions, such as erosion or dilation, to remove small artifacts or smooth
DNNs, on the other hand, learn the segmentation task directly from the segmented regions. Other methods may include region growing,
the data. By optimizing the network parameters using labeled training graph-cut algorithms, or spatial regularization techniques to enforce
data, DNNs can effectively learn to perform the segmentation task in a consistency or incorporate prior knowledge. Post-processing can en-
data-driven manner. This end-to-end learning approach simplifies the hance the segmentation accuracy and improve the clinical usability of
segmentation pipeline and can potentially yield better segmentation the results [29].
results. The U-Net architecture has become popular for medical image
Semantic Segmentation: DNNs, particularly CNNs, are widely used segmentation tasks. It consists of an encoder–decoder structure with
for semantic segmentation in medical imaging. Semantic segmentation skip connections [30]. The encoder path captures context and extracts
aims to assign a label or class to each pixel or voxel in an image, features through convolution and pooling operations, while the decoder
effectively partitioning the image into different regions [26]. CNNs path recovers spatial resolution and refines the segmentation masks.
with fully convolutional architectures can process the entire image at Skip connections allow the network to combine low-level and high-level
once and generate dense pixel-wise predictions. The networks learn to features, aiding in precise localization and preserving fine details [31].
associate image regions with specific classes during training, enabling U-Net and its variations have shown remarkable performance in various
accurate and detailed segmentation. medical imaging segmentation challenges.
Training with Labeled Data: DNNs require large amounts of labeled In this case, the network structure is symmetrical, consisting of
data for training. Annotated medical image datasets are essential for an encoder–decoder framework, and it can be presented as a U-shape
training DNN-based segmentation models [27]. The availability of such Network (U-Net) for 2D images or Volumetric Network (V-Net) [32] for
datasets, along with advancements in data sharing and labeling efforts, 3D MRI images. Below, the conceptual layout of the U-net architecture
has facilitated the development of robust and accurate segmentation depicts in Fig. 3. The set of layers is the same as in standard CNN,
models. However, labeled medical image datasets can still be limited but their organization results in a segmented image, with the original
or require extensive manual annotation, which remains a challenge in resolution as the input one, where each pixel value is represented as
the field [28]. belonging to a particular class [33].
Post-processing Techniques: Post-processing techniques are often Such an architecture is especially useful for medical image segmen-
employed to refine and improve the segmentation results obtained from tation tasks. Here, in many cases, the tissues or organs are difficult to

4
P.K. Mall, P.K. Singh, S. Srivastav et al. Healthcare Analytics 4 (2023) 100216

Fig. 4. An example of 3D organ visualization before an operation by means of a V-Net framework by [35].

Table 1
List of accessible datasets with an access link that exploits DNNs.
Application type Modality Dataset Link
BreastCancer (BC) Digital mammograms DREAM [36] https://www.synapse.org
BC H&E stained BreakHis [37] https://web.inf.ufpr.br/
BC H&E stained MITOSTAPIA [38] https://mitos-atypia-14.grand-challenge.org/
BC H&E stained BACH [39] https://iciar2018-challenge.grand-challenge.org/
BC H&E stained CAMELYON16 [37] https://camelyon16.grand-challenge.org/
BC H&E stained INBreast [40] https://www.ncbi.nlm.nih.gov/pubmed/
BC H&E stained CAMELYON17 [41] https://camelyon17.grand-challenge.org/
Chest X-ray (C X-ray) X-ray CheXperte [42] https://stanfordmlgroup.github.io/competitions/
C X-ray X-ray MIMIC-CXR [43] https://mimic-cxr.mit.edu/
Head And Neck (H & N) CT images PDDCA [44] http://www.imagenglab.com/
H & N PET scans 18F-FDG PET [45] https://www.kaggle.com/c/pet-radiomics-challenges/
Lung Cancer (LC) Low-Dose CT images BOWL17 [46] https://www.kaggle.com/c/data-science-bowl-2017/
LC Computed Tomography (CT) scans LIDC-IDRI [47] https://wiki.cancerimagingarchive.net/
LungMask (LM) CT Scan LUMIC [48] https://lumic.grand-challenge.org/
LungNode (LN) CT Images LUNA [49] https://luna16.grand-challenge.org/
Mice LungLesions H&E stained ANH [50] https://anhir.grand-challenge.org/
Mild traumatic brain injury MRI MTOP [51] https://tbichallenge.wordpress.com/
Multimodal brain tumor segmentation 3T Multimodal MRI scans BraTS 2018 [52] https://www.med.upenn.edu/
Brain 3T MRI Scans MRBrainS18 [53] https://mrbrains18.isi.uu.nl/
Brain CT-MR And PET-MR RIRE [54] http://insight-journal.org/
Brain MRI MUSHAC [55] https://projects.iq.harvard.edu/
Brain tumor MRI Curiou [56] https://curious2018.grand-challenge.org/
Alzheimer’sDiseaseNeuroimagingInitiative MR And PET TADPOLE [57] https://tadpole.grand-challenge.org/
DiabeticRetinopathy FundusImages IDRID [58] https://idrid.grand-challenge.org/
Digital RetinalVessel extraction Non-Mydriatic 3CCD Camera DRIVE [59] https://drive.grand-challenge.org/ [60]
Retinal OCT fluid OCT Scan RETOU [61] https://retouch.grand-challenge.org
CardiacDiagnosis MRI ACDC17 [62] https://www.creatis.insa-lyon.fr/
Cardiac ECG ECG Arrhythmia [63] https://irhythm.github.io/cardiol_test_set/
Pulmonary angiograms CT scans CAD-PE [64] http://www.cad-pe.org
Liver Ultrasound CLUST [65] https://clust.ethz.ch/
Abdominal CT And MRI CHAOS [66] https://chaos.grand-challenge.org/Data/
KidneyCancer CT KITS [67] https://kits19.grand-challenge.org/data/
GastroenterologicalDiseases Endoscopy AIDA [68] https://isbi-aida.grand-challenge.org/
X-ray Normal/Abnormal detection. X-ray MURA [69] https://stanfordmlgroup.github.io/
Low back pain MRI IVDM3S [70] https://ivdm3seg.weebly.com/
Child age prediction X-ray Rsna-bone-age [71] https://www.kaggle.com/
Knee MRI MRNet-v1.0 [72] https://stanfordmlgroup.github.io/
Automatic vertebral fracture analysis and identification X-ray VFA, DXA [73] http://spineweb.digitalimaginggroup.ca/
Tumors H&E stained Warwick-QU [74] https://warwick.ac.uk/
Skin cancer Dermo.S. ISIC [75] https://challenge2019.isic-archive.com/
Skin cancer Dermo.S Skin cancer MNIST [76] https://www.kaggle.com/

recognize in the unprocessed image, even for a professional. Whereas, map and collected multi-scale features, we focus on the boundary of
after segmentation, they become easily noticeable [34]. the target lesions or organs that need to be segmented. The results
The network designers have gone even further and attempted to are then obtained by combining the border and region properties.
segment the diseased organs in 3D for better visualization, especially In [79] authors have suggest Axial Fusion Transformer UNet (AFTer-
when an operation (invasive medical procedure) is necessary. The best UNet), which combines the strengths of transformers for long sequence
example of such an approach is the 3D V-net architecture, which modeling with convolutional layers’ capacity to extract fine-grained
was recently proposed by Ronneberger, Fischer, & Brox in [35]. The data. It uses long-range signals from inside and between slices to direct
authors proposed a 3D liver segmentation and tumor identification segmentation. Compared to the earlier transformer-based models, it
technique based on a 3D V-Net, which is shown in Fig. 4. Here, trains faster and uses fewer parameters. Extensive tests on three multi-
input images came from two datasets, LiTS [77] and 3D-IRCADb (see organ segmentation datasets show that our method performs better
Table 1). In [78] this work, authors have introduce a novel region- than the most recent state-of-the-art approaches.
to-boundary deep learning model to provide a practical solution to This architecture has many variations, including the DENSE-
this problem. Authors have initially use a U-shaped network with two INception U-net, among others [80]. This solution has been tested on
branches below the final layer to build the target probability map and different organ images, for instance, lungs segmentation of CT data
the associated signed distance map. Second, using the signed distance obtaining an average Dice score of 98.57%, blood vessel segmentation

5
P.K. Mall, P.K. Singh, S. Srivastav et al. Healthcare Analytics 4 (2023) 100216

obtaining an average Dice score of 95.82%, and MRI scan brain tumor
segmentation with an average Dice-score 98.67%.

2.3. DNN architecture for object detection

Deep neural networks (DNNs) have had a profound impact on

object detection, which involves locating and classifying objects within
images or videos. The significance of DNNs for object detection can be
understood through the following key points:
Accuracy: DNNs have demonstrated exceptional accuracy in ob-
ject detection tasks. Through the use of deep architectures, such as
convolutional neural networks (CNNs), DNNs can learn complex and
discriminative features from images, enabling precise object localiza- Fig. 5. The conceptual layout of Faster R-CNN architecture [88].
tion and classification [81]. The hierarchical nature of DNNs allows
for the extraction of features at multiple scales and levels of abstrac-
tion, capturing both local and global information. This leads to highly rich contextual information. These advancements have made DNN-
accurate object detection results, outperforming traditional methods. based object detection a critical technology in various fields, includ-
End-to-End Learning: DNNs facilitate end-to-end learning, where the ing computer vision, autonomous systems, robotics, surveillance, and
entire object detection pipeline can be learned directly from the data. beyond.
Traditional object detection methods often involve multiple stages, The GenPSOWVQ approach [86], a novel wavelet-based recurrent
including feature extraction, region proposal generation, and classi- neural network technique for image reduction. Combining genetic algo-
fication. DNNs, on the other hand, can learn to perform all these rithms and fragments, the codebook is created. When encoding clinical
steps simultaneously within a single network. This end-to-end learning pictures, the recently developed image compression model achieves
approach reduces complexity, eliminates the need for handcrafted fea- exact compression while retaining image accuracy with lower comput-
tures, and optimizes the entire detection process, resulting in improved ing costs. Real-time medical imaging was used to test the suggested
accuracy and efficiency [82]. technique utilizing PSNR, MSE, SSIM, NMSE, SNR, and CR indica-
Flexibility and Adaptability: DNNs offer flexibility and adaptability tors. [87] author have suggest the Mixed Transformer Module (MTM), a
to various object detection scenarios. They can detect objects across brand-new Transformer module for simultaneous learning of intra- and
different classes, sizes, orientations, and scales. The training process inter-affinities. First, MTM efficiently determines self-affinities using
of DNNs involves optimizing the network parameters using labeled our clever Local–Global Gaussian-Weighted Self-Attention (LGG-SA).
datasets, allowing them to learn the specific characteristics of objects Then, it uses External Attention (EA) to mine links between data
within a given domain. This adaptability makes DNNs suitable for a samples. We create the Mixed Transformer U-Net (MT-UNet) model
wide range of applications, from detecting everyday objects in images for precise medical picture segmentation using MTM. The experimental
to specialized tasks like detecting specific anatomical structures in findings demonstrate that the suggested technique outperforms existing
medical imaging [83]. state-of-the-art methods on two separate public datasets where tested.
Real-Time Performance: DNNs have made real-time object detection Last but not least, DNN architecture can be used for object de-
possible. With advancements in hardware, optimization techniques, tection. In this case, two network functions are used: classification
and spatial location. These functions can be obtained thanks to the
and model architectures, DNN-based object detection can achieve
Residual Networks, consisting of many similar modules, e.g., RezNet-50
high frame rates, enabling applications in real-time video analysis,
or RezNet-101 [88]. One such module is illustrated in Fig. 5. The main
autonomous driving, surveillance, and robotics. Techniques like single-
flow is responsible for object classification, whereas the feedforward is
shot detectors (SSDs) and region-based convolutional neural networks
responsible for the transfer of spatial location information.
(R-CNNs) have been developed to balance accuracy and speed, making
In medical imaging, R-CNN can be used to detect objects, for
DNN-based object detection practical for real-world applications [84].
instance, in X-ray images [89]. This architecture can even be ap-
Transfer Learning and Pre-trained Models: DNNs facilitate transfer
plied in a real-time segmentation technique, which is presented by the
learning, where models trained on large-scale datasets can be fine-
authors [90], who used it in a prostate biopsy procedure.
tuned for specific object detection tasks with limited training data.
We propose anEnhance-NetDLM approach to improve the perfor-
Pre-trained models, trained on massive datasets like ImageNet, serve as
mance of musculoskeletal radiographs X-ray images. The highlight of
a starting point, allowing users to leverage the knowledge and feature this research is to assess the impact of three different image enhance-
representations learned from these datasets. This significantly reduces ment techniques (CLAHE, HEF, and UM) on green channel grayscale
the amount of data and computation required to train object detection the medical musculoskeletal radiographs X-ray images for DLMs. The
models from scratch, making it more accessible and feasible for various paper is organized into six different sections as follows: the most crucial
applications [85]. related works are described in the second section. In the third section,
Rich Contextual Information: DNNs can effectively utilize rich con- we have discussed the materials and methods used for the proposed
textual information for object detection. By learning from large-scale model. In the fourth section, we have elaborated on the proposed
datasets, DNNs can capture and understand the contextual relationships model. In the fifth section, we have explained the simulation, results,
between objects and their surrounding environment. This contextual and validation in detail. In the last section, we have discussed the
understanding improves the accuracy of object detection by incorpo- conclusion and future scope.
rating contextual cues, spatial relationships, and semantic information.
For example, DNNs can exploit the context of object co-occurrences, 3. Medical image datasets organized according to organs
object shapes, or scene context to improve the detection and reduce
false positives. Neural networks (regardless of their architecture) need data for
DNNs have revolutionized object detection by significantly improv- training. Let us summarize most of the publicly available medical
ing accuracy, enabling end-to-end learning, offering flexibility, achiev- imaging data sets according to the pattern recognition tasks in the
ing real-time performance, supporting transfer learning, and leveraging context in which they have been applied. The process of neural training

6
P.K. Mall, P.K. Singh, S. Srivastav et al. Healthcare Analytics 4 (2023) 100216

requires a big training set. In the case of medical imaging, several In some cases, the image registration turned out to be a problem.
problems arise frequently; therefore, an important contribution of this Berendsen et al. [93] implemented a ConvNet-based model for 3D
article is putting together and subjecting to critical analysis the medical image registration, and after limited training, the average registration
image sets divided into organs and image modalities. The research time was as good as 0.63 s for multistage image registration.
was carried out in different anatomical regions to support medical Ferreira et al. [96] suggested a V-Net model for segmentation tasks
practitioners. Table 1 lists accessible datasets that can be used for DNN that was fully regularized. It implements regularization techniques,
training. In addition to the link to the actual data set, information about such as dropout, multi-task learning, deep-supervision, and batch nor-
the specific modality that is represented is provided. This article [91] malization. The model performs an average per-lobe Dice Coefficient
covers a machine learning and image processing-based evolutionary of 93.6% and an inter-lobar Dice Coefficient of 76.2%.
method for categorizing and identifying breast cancer. To assist in The outbreak of the COVID-19 pandemic resulted in an urgent
the classification and detection of skin disorders, this model integrates necessity for fast respiratory tract diagnosis in case of pneumonia and
image preprocessing, feature extraction, feature selection, and machine other lung infections. Oh, Park and Ye [97] implemented a combination
learning approaches. The geometric mean filter is applied to improve of FC-DenseNet-103 for segmentation and ResNet-18 for classification.
the quality of the image. Features are extracted using AlexNet. The An accuracy rate of 88.9 percent was reached by the model. Ucaret
relief algorithm is used to pick features. The model uses machine al. [98] suggested a deep Bayes-Squeeze Net model (convolutional
learning methods such least square support vector machine, KNN, network with Bayesian optimization) diagnose COVID-19. The pro-
random forest, and Nave Bayes for illness classification and detection. posed approach solves the dataset imbalance issue and implements a
Data gathering from MIAS is utilized in the experimental inquiry. Using multiscale offline augmentation. The experiment reported shows an
image analysis to correctly diagnose breast cancer illness is a benefit of accuracy of 98.26%.
the suggested method. Wu et al. [99] proposed a deep multi-view deep learning fu-
The collected datasets were selected by clinical experts and manu- sion model to diagnose COVID-19. Based on the modified ResNet-50
ally prepared from the most interesting point of medical view. and the experiment shows an accuracy of 76.0%. Behzadi-Khormouji
Databases include several sets of anonymized medical images of pa- et al. [100] implemented a three-phase pre-processing method to
tients. Some of the above datasets have been goals of different chal- improve the DNN model, referred to as ChestNet. The experiment
lenges organized to evaluate new and existing algorithms for automated achieved an accuracy of 94.67%. Mittal et al. [101] implemented
classification, detection, or segmentation in order to reduce the time two models for the integration of convolutions with capsules (ICC)
required to diagnose a patient. and an ensemble of convolutions with capsules (ECC) and achieved
Transfer Learning (TL) architectures for automated medical image an accuracy of 95.33% and 95.90%, respectively. As defined by the
processing. Authors have revealed that TL has been used for a variety of Neuronal Activity Vector (NAV), a capsule is a small collection of
medical imaging applications, including segmentation, object recogni- neurons whose activity vector indicates the precise parameters of an
tion, illness classification, and severity grading, to mention a few [92]. entity of a given kind, such as a component of an item or the object
We could prove that TL uses less training data than conventional deep
itself, introduced by Sabour, Frosst, & Hinton [102].
learning techniques while still offering high-quality decision assistance.
Chest X-rays are used to see how much virus is in the lungs in differ-
The fact that TL models have previously been trained on sizable generic
ent sections of the body. Chest radiographs are the preferred method of
datasets and that a task-specific dataset is merely utilized to customize
diagnosis since laboratory testing takes a long time and requires a great
the model gives birth to these favorable qualities.
deal of human work. It is possible to determine if a patient is suffering
For a medical purpose, information not only from pure images
from viral-pneumonia, COVID-19, or has a normal profile by looking
or one modality is used. Additionally, the process of image registra-
at their chest X-ray scans [103]. For binary classification, their model
tion, which is aligning two or more images, has had to be employed.
was 99.48% accurate, while for three-class classification, it was 97%
In (semi-)automatic medical image analysis, it is a well-established
accurate.
approach for transferring information across pictures. Methods for
Overall, researchers have successfully trained different DNN archi-
registering images based on picture intensity and characteristics that
tectures for the classification, detection, registration, and segmentation
are manually constructed are among the most common. Images may
of lung disease with high accuracy [104]. Table 2 summarizes the
now be registered with the use of supervised and unsupervised deep
above-mentioned approaches and results in the most compact way.
learning algorithms in recent years [93,94].

4. DNN applications using medical imaging datasets 4.2. DNN applications for eye diseases

With knowledge of main NN architectures and popular datasets DNN has also been applied in ophthalmology at the same time as
that can be used for their training, let us now summarize recent for the diagnostics of other organs, e.g., to retinal fundus images for
developments and research work carried out in the different anatomical the classification of diabetic retinopathy; predicting age and gender,
regions. retinopathy, macular edema, glaucoma-like disk. Nagasato et al. [105]
The primary goal of this part is to demonstrate how and where implemented the Visual Geometry Group VGG-16 model [106]. The
publicly accessible medical imaging data sets are utilized in the area analysis showed an AUC of 98.6%.
of DNN architectures described in Section 2. The research was carried Wu, Xia et al. [107] proposed the NFN+ model, which contains two
out in the different anatomical regions, such as the kidney, liver, brain, cascaded, identical multiscale structures for preparing a retinal vessel
breast, chest, eye, etc., to support medical practitioners. map. Each segmented network framework has an encoder–decoder
architecture with 3 × 3 Conv-ReLU layers (compare Fig. 3). The model
4.1. DNNapplications in lung diseases achieved AUC for three different datasets, DRIVE, STARE, and CHASE
(see Table 1 for characteristics of the DRIVE dataset) of 98.30%,
In this subsection, we present some examples of the above- 98.75%, and 98.94%, respectively.
mentioned DNN techniques in lung diseases. Before the COVID-19 Riaz et al. [108] proposed a multiclass, deep, and densely connected
pandemic, the first attempts were made for diagnoses lung diseases model. There, the EyePACS dataset was used for the experiment. The
based on CT images. Abdelhamid et al. [95] implemented the U-Net analysis showed an average F1score of 97.0%. Liu et al. [109] im-
model, trained as an end-to-end model with a performance of 95.02% plemented a self-adaptive deep learning method based on Inception-
of the Dice-Coefficient index. V3, and the analysis showed an AUC of 99.50%. Zhang et al. [80]

7
P.K. Mall, P.K. Singh, S. Srivastav et al. Healthcare Analytics 4 (2023) 100216

Table 2
Summary of DNN techniques for lung diseases.
Reference Studied Task Modality Architecture/Layers Result
Oh Park and Ye [97] COVID-19 Detection X-ray Combination of FC-DenseNet103 Accuracy 88.9%
and ResNet18
Ucar and Korkmaz [98] COVID-19 Detection X-ray Deep Bayes-SqueezeNet Accuracy 98.26%
Wu et al. [99] COVID-19 Detection CT Multi-view deep learning fusion model Accuracy, 76.0%
Abdelhamid et al. [95] Lung Segmentation X-ray Unet Model performance 0.9502
Dice-Coefficient index
Berendsen et al. [93] Lung Registration X-ray Deep Learning Image Registration The average registration time for
multistage image registration was 0.63 s.
Ferreira et al. [96] LungLobe Segmentation 3D CT V-Net The model performs an il-DC Coefficient
of 76.2% and an average pl-DC of
93.6%.
Mittal et al. [101] Pneumonia Detection X-ray ICC and ECC Achieved accuracy of 95.33% and
95.90%, respectively
Behzadi-Khormouji et al. [100] Respiratory disease Detection X-ray ChestNet Model accuracy 94.67%

Table 3
Summary of DNN techniques for eye diseases.
Reference Studied Application Architecture/Layers Modality Result
Nagasato et al. Detect a nonperfusion area (NPA) Classification VGG-16 model Optical coherence AUC 98.6%
[105] tomography angiography
(OCTA) images
Wu, Xia et al. DiabeticRetinopathy Segmentation NFN+ Model FundusRetinal image AUC DRIVE, STARE, CHASE
[107] 98.30%, 98.75%, 98.94%
respectively
Riaz et al. [108] DiabeticRetinopathy Classification Multiclass deep and densely Fundus-Retinal image Average F1-score AUC97.0%.
connected mode
Zhang et al. [80] Blood vessel segmentation, Lung’s Segmentation Seg-Net, FCN-8s, U-Net, U-Net, CT, Fundus-Retinal image, Average-Dice-Score of 0.9867
segmentation, and Brain tumor DIU-Net-1, DIU-Net MRI
segmentation

model combine the inception unit with dense connections in the U-Net S. S. Halabi et al. [71] performed a study concerning pediatric
framework. The experiment showed an Average Dice score of 98.67%. boneage based on 14 236 images and implemented Inception V3,
Kim et al. [110] implemented the CNN model trained using retinal ResNet-50, Ice Module Architecture, ML using handcrafted features and
fundus images age prediction. The model achieved MAE of 3.06 on the Gabor texture energies, U-Net. The analysis results, based on a mean
test set of 24,366 images. absolute distance (MAD), revealed 4.2, 4.4, 4.4, 4.5, and 4.5 months,
The summary of the DNN technique for ophthalmology classifi- respectively.
cation, segmentation, detection, and prediction using different image S. Koitka et al. [114] implemented an Inception-ResNet-V2 model.
modalities is shown in Table 3. Overall, these researches cover the The analysis showed the average F1-Score to be 91.85%.
retinopathy resulting from different causes and diagnosed based on S. H. Tajmir et al. [115] focused on building a CNN model based on
various imaging. LeNet-5 introduced in 1989 by Yann LeCun [116], with a stride of 4 and
image patch size 32 × 32, for bone age assessment (BAA) as a tool in the
4.3. DNN applications for bone age diseases
calculation of metabolic disorders and pediatric endocrine. Bone-age
assessment performance was found to be 68.2 percent accurate overall
The skeleton or bone age automatic evaluation is a very new subject
and 98.6 percent within a year of the model.
for deep learning. So far, visual assessment of boneage, based on a
V. I. Iglovikov et al. [117] performed a study on 12,600 radiological
solitary reference X-ray image, has been the standard, and due to the
images and implemented VGG-style DNN 6 convolutional layers fol-
fact that such assessment is a time-limited process, it is subject to the
lowed by two fully connected blocks. There was a mean absolute error
variability of interpretations.
of 4.97 months in the procedure.J. C. Castillo et al. [118] performed the
Most of the ongoing research focuses on designing a novel approach
study on 14 949 radiological images and implemented VGG16. Patients
to predict boneage, using left-hand radiology images for a child. C.
Spampinato et al. [111] proposed a two-stage DNN model, a first- of both genders, male and female, had an MAE of 9.82/10.75 months
stage CNN network for automatic feature extraction, and a second-stage when using the CNN design.
regression model for bone-age prediction. They obtained an average T. Van Steenkiste et al. [119] performed the study based on 12 611
evaluation discrepancy of about 0.8 years. radiograph images of children from the age group of 0 to 19 years. The
H. Lee et al. [112] performed a study based on 8325 images Gaussian process regression was applied to aggregate results obtained
and proposed a data-driven CNN transfer learning for automatic fea- from the augmented DNN model, which yielded an accuracy rate of
ture extraction. Their model has accomplished an accuracy of 57.32% 94.45% and a mean spread of 6.80 months inside a year.
and 61.40% for both gender, females and male, respectively. Larson Transfer learning was examined by Kandel et al. [120] for its impact
et al. [113] performed a study based on 14 036 images and imple- on the classification of musculoskeletal pictures. After analyzing the
mented ResNet-50 architecture. The analysis showed validation and 168 results acquired utilizing six different architectures for CNNs and
test accuracies of 66.2% and 49.7%, respectively. seven distinct kinds of bones, the scientists discovered that the transfer

8
P.K. Mall, P.K. Singh, S. Srivastav et al. Healthcare Analytics 4 (2023) 100216

Table 4
Summary of DNN techniques for bone age prediction using X-ray image modality.
Reference Dataset size Architecture/Layers Result
C. Spampinato et al. [111] 1391 BoNet architecture. The average discrepancy is about 0.8 years.
H. Lee et al. [121] 8325 CNN-based pre-processing engine Fine-tuned CNN accuracies of 57.32% and 61.40%
images. for females and males.
D. Larson et al. [114] 14 236 ResNet-50 Testand validation accuracies are 0.497 and 0.662,
respectively.
S. S. Halabi et al. [71] 14 236 Inception V3, ResNet- 50, Ice Module Using MAD, the best five outcomes were 4.2, 4.4,
Architecture, ML using handcrafted features 4.4, 4.5, and 4.5 months, respectively.
and Gabor texture energies, U-Net.
S. Koitka et al. [114] 14 236 Inception-ResNet-V2 F1-Score on average, 91.85%.
S. H. Tajmir et al. [115] 1681 LeNet-5 According to the results mean six-reader cohort
accuracy was 63.6 percent overall and 97.4
percent within one year.
V. I. Iglovikov et al. [117] 12 600 VGG-style CNN The method achieved a mean absolute error (MAE)
equal to 4.97 months.
J. C. Castillo et al. [118] 14 949 VGG-16 Patients of both genders, male and female, had an
MAE of 9.82/10.75 months when using the CNN
design.
T. Van Steenkiste et al. [122] 14 236 VGG-16 The method has a 94.45 percent accuracy rate and
a mean range of 6.80 months after one year.
Kandel et al. [120] Training: normal 21,935 fractured Wrist: DenseNet, Finger: InceptionResNetV2, The overall accuracy of different models oscillated
14,873 Elbow: Xception, Forearm: DenseNet121, from 70% to 80%, which is compared to
Test: normal 1667 fractured 1530 Humerus: VGG-19, Hand: Xception, radiologists’ results.
Shoulder: Xception.

learning approach outperformed training a CNN from scratch in terms study, we propose a hybrid artificial intelligence (AI) framework to
of accuracy. identify major chronic risk factors of novel, contagious diseases as
The summary of the DNN technique for bone age prediction using early as possible at the time of pandemics. The proposed framework
X-ray image modality is shown in Table 4. Our summary covers the combines evolutionary search algorithms with machine learning and
research concerning bone age for all age groups, from children to old the novel explanatory AI (XAI) methods to detect the most critical
people. risk factors, use them to predict patients at high risk of mortality, and
analyze the risk factors at the individual level for each high-risk patient.
4.4. DNN applications for osteosarcoma diagnostics H. B. Arunachalam et al. [127] focused on fine-tuning hyper pa-
rameters, based on gradient descent, with a learning rate of 10-3 and a
Digital automation in the area of histopathology is restricted be- batch size of 100. The model was run for 20 epochs and achieved 93%
cause of the intricacy of tissue structure. Osteoblasts, osteocytes, and of accuracy. M. D Acunto et al. [128] implemented a Faster R- CNN
osteoclasts serve as a visual representation of the various tissue types model. The analysis showed an accuracy of 97%.
seen in a typical bone biopsy sample. The malignant characteristics The summary of the DNN technique for osteosarcoma cell detection
of these cells include nuclear membrane abnormalities, pleomorphism, and classification using H&E image modality is shown in Table 5.
large multinucleated cells, hyperchromatic nuclei, and aberrant mitosis. Because of structural complexity and difficulties in recognition between
Hematoxylin and eosin (H&E) stained datasets of cancer cells on normal and disease cells, there have been many architectures tested
a histology slide are used in the majority of cancer research. H&E with relatively small size data sets.
staining dyes the nuclei blue and the background tissues pink, and
it is used in the majority of cancer research. The slide preparation 5. ChallengesandFutureDirections
and incorrect staining reaction have an impact on the digital picture
quality, which may result in numerous cellular areas and tissues being This paper summarizes how deep neural networks can help the
under-represented in the digital image. doctor determine a clinical situation. As it has been shown, DNN-
R. Mishra et al. [123] designed a CNN, which concentrated on a based approaches produce low false-positive levels and speed up the
training phase of categorization into three classes: live tumor, necrosis, decision process. It has been demonstrated that different deep learning
and non-tumor. Assigning tumor classifications is accomplished by algorithms offer greater accuracy in the field of medical imaging than
using the trained classifier as a sliding window on top of the collected human expert evaluation.
data. Because CNN works on raw pixel data, there is no need for The remote hospital, especially in the COVID-19 era, should offer
human input. A noteworthy benefit over earlier efforts, apart from the considerable assistance not only to patients but also to doctors. In order
initial annotation of slides for the training data, was the attainment of to achieve the aim of health for all, virtual hospitals should offer the
classification accuracy of roughly 84 percent, which was a substantial same degree of diagnostic sophistication to an individual in a rural
improvement over past attempts. community as well as to an urban one, generally independently of
O. Daescu et al. [124] performed a study on 64,000 osteosarcoma the geographic location of the patient. In the future, virtual hospitals,
image patches, resized to 128 × 128, and proposed a CNN model, thanks to a robust computer infrastructure applying DNNs to image
which achieved a classification accuracy of around 92.4%. In [125] processing, will help to build a remote healthcare system.
description of explainable artificial intelligence (XAI), which is applied Representative datasets have played a critical role in the advance-
to medical picture analysis using deep learning. For the purpose of ment of research and innovation in a wide range of high-profile areas of
categorizing deep learning-based medical image analysis techniques, a data science, including computer vision. In the process of monitoring
framework of XAI criteria is presented. The papers are then surveyed an ongoing disease, the development of medical imaging is essential.
and categorized in accordance with the framework and according to MRI, CT scans, and other medical modalities allow the doctor to
anatomical location for use in medical image analysis. In [126] this track the treatment effectiveness and change procedures where it is

9
P.K. Mall, P.K. Singh, S. Srivastav et al. Healthcare Analytics 4 (2023) 100216

Table 5
Summary of DNN technique for osteosarcoma cell detection and classification using H&E-stained image modality.
Reference Osteosarcoma studied Dataset size Architecture/Techniques Result
R. Mishra et al. [123] Viable and non-viable tumor 5000 AlexNet, LeNet, VGGNet, and Proposed AlexNet: 0.73, LeNet: 0.67, and Proposed
prediction architecture architecture: 0.84
O. Daescu et al. [127] Histopathological analysis of 64,000 AlexNet, LeNet, VGGNet, Baseline The overall accuracy of different models
osteosarcoma architecture, and Proposed architecture AlexNet: 0.78, LeNet: 0.71, and Proposed
architecture: 0.867
H. B. Arunachalam et al. [129] Human osteosarcoma cell detection 942 Decision trees, SVM, Ensemble learning, The overall accuracy of different techniques:
and classification and CNN Decision trees, SVM, Ensemble learning, and
CNN achieves 80.9%, 89.9%, 86.6%, and
93%.
M. DAcunto et al. [128] Human osteosarcoma cell detection 48 Faster R-CNN The accuracy is 0.97.
and classification

appropriate. The detailed knowledge provided by 3D medical imaging References

offers more thorough treatment for the patients. This high precision
obtained thanks to 3D CNNs is used for medical image diagnosis. [1] J.L. McClelland, D.E. Rumelhart, P.D.P.R. Group, et al., Parallel Distributed
Processing, Vol. 2, MIT Press, Cambridge, MA, 1986.
[2] S. Suganyadevi, V. Seethalakshmi, K. Balasamy, A review on deep learning in
6. Conclusion medical image analysis, Int. J. Multimed. Inf. Retr. 11 (1) (2022) 19–38.
[3] V. Chang, V.R. Bhavani, A.Q. Xu, M.A. Hossain, An artificial intelligence model
This article presents a comprehensive overview of the recent devel- for heart disease detection using machine learning algorithms, Healthc. Anal.
2 (2022) 100016.
opments in the area of deep learning concerning the diagnosis process.
[4] M. Izadikhah, A fuzzy stochastic slacks-based data envelopment analysis model
It focuses on organ diseases, such as kidney, liver, breast, brain, chest, with application to healthcare efficiency, Healthc. Anal. 2 (2022) 100038.
eye, bone fracture, bone age prediction, and osteosarcoma based on [5] M.K. Rohil, V. Magotra, An exploratory study of automatic text summarization
medical imaging. It covers four primary directions. in biomedical and healthcare domain, Healthc. Anal. 2 (2022) 100058.
In the first section, the background to DNN development in medical [6] M.S. Pathan, A. Nag, M.M. Pathan, S. Dev, Analyzing the impact of feature
selection on the accuracy of heart disease prediction, Healthc. Anal. 2 (2022)
imaging is outlined. The second section provides an overview of the
100060.
basic DNN architectures and a case study of the top DNN frameworks. [7] G. Huang, Y. Sun, Z. Liu, D. Sedra, K.Q. Weinberger, Deep networks with
Ongoing research in this sector, such as CNNs, has been successfully stochastic depth, in: European Conference on Computer Vision, 2016, pp.
implemented in medical image classification, V-Nets in segmentation, 646–661.
and ResNet in detection. The 3D V-Nets are complex neural systems [8] C. Nwankpa, W. Ijomah, A. Gachagan, S. Marshall, Activation functions:
Comparison of trends in practice and research for deep learning, 2018, pp.
used in recent times to examine clinical data. The third section elabo- 1–20.
rates on article contributions to the open-source datasets which can be [9] M.F. Kabir, T. Chen, S.A. Ludwig, A performance analysis of dimensionality
used in medical image analysis. reduction algorithms in machine learning models for cancer prediction, Healthc.
In the fourth section, the article focuses on the research work done Anal. 3 (2023) 100125.
[10] F. Masood, M. Driss, W. Boulila, J. Ahmad, S.U. Rehman, S.U. Jan, A. Qayyum,
in the field of deep learning in order to assist the healthcare worker
W.J. Buchanan, A lightweight chaos-based medical image encryption scheme
in offering treatment for diseases. Materials are divided according to using random shuffling and XOR operations, Wirel. Pers. Commun. 127 (2)
the organs. The authors have discussed multiple deep learning frame- (2022) 1405–1432.
works for medical imaging and the prospects for their applications, [11] B. Davazdahemami, P. Peng, D. Delen, A deep learning approach for predicting
especially for the machine learning and computer vision community, early bounce-backs to the emergency departments, Healthc. Anal. 2 (2022)
100018.
to solve medical imaging tasks. The goal of a lot of ongoing research
[12] Y. Chen, X.-H. Yang, Z. Wei, A.A. Heidari, N. Zheng, Z. Li, H. Chen, H.
is to aid medical professionals by developing automated systems to Hu, Q. Zhou, Q. Guan, Generative adversarial networks in medical image
examine images and identify acute disorders including brain tumors, augmentation: A review, Comput. Biol. Med. 144 (2022) 105382.
bone cancer, breast cancer, bone fracture, and many more. This thor- [13] Y. Skandarani, P.-M. Jodoin, A. Lalande, GANs for medical image synthesis: An
empirical study, J. Imaging 9 (3) (2023).
ough analysis provides a summary of current developments in deep
[14] J. Cheng, S. Tian, L. Yu, C. Gao, X. Kang, X. Ma, W. Wu, S. Liu, H. Lu,
neural network-based medical imaging. In addition to the thorough ResGANet: Residual group attention network for medical image classification
literature evaluation, a list of publicly accessible data sources and and segmentation, Med. Image Anal. 76 (2022) 102313.
potential future study areas are outlined. In the future, medical com- [15] Q. Guan, Y. Chen, Z. Wei, A.A. Heidari, H. Hu, X.-H. Yang, J. Zheng, Q.
munities and patients will benefit from addressing challenges related Zhou, H. Chen, F. Chen, Medical image augmentation for lesion detection using
a texture-constrained multichannel progressive GAN, Comput. Biol. Med. 145
to diagnosing critical diseases. Deep learning models can analyze med-
(2022) 105444.
ical images in conjunction with other clinical data, such as genomics [16] X. Chen, X. Wang, K. Zhang, K.-M. Fung, T.C. Thai, K. Moore, R.S. Mannel,
or electronic health records, to enable personalized medicine. By in- H. Liu, B. Zheng, Y. Qiu, Recent advances and clinical applications of deep
tegrating patient-specific information, DNNs can assist in treatment learning in medical image analysis, Med. Image Anal. 79 (2022) 102444.
planning, prognostication, and selection of optimal therapies based on [17] K. Fukushima, Neocognitron: A hierarchical neural network capable of visual
pattern recognition, Neural Netw. 1 (2) (1988) 119–130.
image-derived features and patterns. [18] A. Krizhevsky, I. Sutskever, G.E. Hinton, Imagenet classification with deep
convolutional neural networks, Adv. Neural Inf. Process. Syst. 25 (2012)
Declaration of competing interest 1097–1105.
[19] O. Ronneberger, P. Fischer, T. Brox, U-net: Convolutional networks for
biomedical image segmentation, in: International Conference on Medical Image
The authors declare that they have no known competing finan- Computing and Computer-Assisted Intervention, 2015, pp. 234–241.
cial interests or personal relationships that could have appeared to [20] P. Malhotra, S. Gupta, D. Koundal, A. Zaguia, W. Enbeyle, Deep neural networks
influence the work reported in this paper. for medical image segmentation, J. Healthc. Eng. 2022 (2022) 9580991.
[21] R. Ali, R.C. Hardie, B.N. Narayanan, T.M. Kebede, IMNets: Deep learning
using an incremental modular network synthesis approach for medical imaging
Data availability
applications, Appl. Sci. 12 (11) (2022).
[22] X. Xie, X. Pan, W. Zhang, J. An, A context hierarchical integrated network for
Data will be made available on request. medical image segmentation, Comput. Electr. Eng. 101 (2022) 108029.

10
P.K. Mall, P.K. Singh, S. Srivastav et al. Healthcare Analytics 4 (2023) 100216

[23] P.R. Jeyaraj, E.R.S. Nadar, Medical image annotation and classification employ- [51] Mild traumatic brain injury outcome prediction, 2021, [Online]. Available:
ing pyramidal feature specific lightweight deep convolution neural network, https://tbichallenge.wordpress.com/data/. [Accessed: 11-Dec-2021].
Comput. Methods Biomech. Biomed. Eng. Imaging Vis. (2023) 1–12. [52] B.H. Menze, A. Jakab, S. Bauer, J. Kalpathy-Cramer, K. Farahani, J. Kirby, Y.
[24] T. Zhou, L. Li, G. Bredell, J. Li, J. Unkelbach, E. Konukoglu, Volumetric memory Burren, N. Porz, J. Slotboom, R. Wiest, et al., The multimodal brain tumor
network for interactive medical image segmentation, Med. Image Anal. 83 image segmentation benchmark (BRATS), IEEE Trans. Med. Imaging 34 (10)
(2023) 102599. (2014) 1993–2024.
[25] L. Xie, L.E.M. Wisse, J. Wang, S. Ravikumar, P. Khandelwal, T. Glenn, A. [53] Grand challenge on MR brain segmentation at MICCAI 2018, 2021, [Online].
Luther, S. Lim, D.A. Wolk, P.A. Yushkevich, Deep label fusion: A generalizable Available: https://mrbrains18.isi.uu.nl/data/. [Accessed: 13-Dec-2021].
hybrid multi-atlas and deep convolutional neural network for medical image [54] RIRE, 2020, [Online]. Available: http://insight-journal.org/rire/index.php.
segmentation, Med. Image Anal. 83 (2023) 102683. [Accessed: 14-May-2020].
[26] X. Xie, W. Zhang, X. Pan, L. Xie, F. Shao, W. Zhao, J. An, CANet: Context aware [55] Multi-modality, lower spine 3D, 2020, [Online]. Available: https://projects.iq.
network with dual-stream pyramid for medical image segmentation, Biomed. harvard.edu/cdmri2018/data. [Accessed: 11-May-2020].
Signal Process. Control 81 (2023) 104437. [56] Curious2018.grand-challenge, 2021, [Online]. Available: https://curious2018.
[27] L. Zhang, K. Zhang, H. Pan, SUNet++: A deep network with channel attention grand-challenge.org/Data/. [Accessed: 11-Sep-2021].
for small-scale object segmentation on 3D medical images, Tsinghua Sci. [57] Tadpole.grand-challenge, 2021, [Online]. Available: https://tadpole.
Technol. 28 (4) (2023) 628–638. grandchallenge.org. [Accessed: 27-Sep-2021].
[28] B. Zhan, E. Song, H. Liu, Z. Gong, G. Ma, C.-C. Hung, CFNet: A medical image [58] idrid.grand-challenge.org. [Online]. Available: https://idrid.grand-challenge.
segmentation method using the multi-view attention mechanism and adaptive org/Data.
fusion strategy, Biomed. Signal Process. Control 79 (2023) 104112. [59] J. Staal, M.D. Abràmoff, M. Niemeijer, M.A. Viergever, B. Van Ginneken, Ridge-
[29] Y. Zhang, R. Jiao, Q. Liao, D. Li, J. Zhang, Uncertainty-guided mutual based vessel segmentation in color images of the retina, IEEE Trans. Med.
consistency learning for semi-supervised medical image segmentation, Artif. Imaging 23 (4) (2004) 501–509.
Intell. Med. 138 (2023) 102476. [60] DRIVE, 2021, [Online]. Available: https://drive.grand-challenge.org/. [Ac-
[30] A. Deshpande, A novel approach to enhance effectiveness of image segmentation cessed: 30-Sep-2021].
techniques on extremely noisy medical images BT - recent trends in image [61] Retouch.grand-challenge, 2021, [Online]. Available: https://retouch.
processing and pattern recognition, 2023, pp. 91–119. grandchallenge.org. [Accessed: 11-May-2021].
[31] Q. He, Q. Yang, M. Xie, HCTNet: A hybrid CNN-transformer network for breast [62] Automated cardiac diagnosis challenge -2017, 2021, [Online]. Available: https:
ultrasound image segmentation, Comput. Biol. Med. 155 (2023) 106629. //www.creatis.insa-lyon.fr/Challenge/acdc/databases.html. [Accessed: 20-Sep-
[32] F. Milletari, N. Navab, S.-A. Ahmadi, V-net: Fully convolutional neural networks 2021].
[63] ECG rhythm classification, 2021, [Online]. Available: https://irhythm.github.
for volumetric medical image segmentation, in: 2016 Fourth International
io/cardiol_test_set/. [Accessed: 23-May-2021].
Conference on 3D Vision (3DV), 2016, pp. 565–571.
[64] Computed tomography pulmonary angiograms CTPA, 2020, [Online]. Available:
[33] P. Tang, P. Yang, D. Nie, X. Wu, J. Zhou, Y. Wang, Unified medical image seg-
http://www.cad-pe.org/?page_id=12. [Accessed: 11-May-2020].
mentation by learning from uncertainty in an end-to-end manner, Knowl.-Based
[65] Challenge on liver ultrasound tracking, 2021, [Online]. Available: https://clust.
Syst. 241 (2022) 108215.
ethz.ch/. [Accessed: 15-Aug-2021].
[34] P. Ashtari, D.M. Sima, L. De Lathauwer, D. Sappey-Marinier, F. Maes, S.
[66] A.E. Kavur, N.S. Gezer, M. Barış, S. Aslan, P.-H. Conze, V. Groza, D.D. Pham,
Van Huffel, Factorizer: A scalable interpretable approach to context modeling
S. Chatterjee, P. Ernst, S. Özkan, et al., CHAOS challenge-combined (CT-MR)
for medical image segmentation, Med. Image Anal. 84 (2023) 102706.
healthy abdominal organ segmentation, Med. Image Anal. 69 (2021) 101950.
[35] Y. Zhang, X. Pan, C. Li, T. Wu, 3D liver and tumor segmentation with CNNs
[67] N. Heller, N. Sathianathen, A. Kalapara, E. Walczak, K. Moore, H. Kaluzniak, J.
based on region and distance metrics, Appl. Sci. 10 (11) (2020) 3794.
Rosenberg, P. Blake, Z. Rengel, M. Oestreich, J. Dean, M. Tradewell, A. Shah,
[36] DREAM, The Digital Mammography DREAM Challenge, DREAM, 2016,
R. Tejpaul, Z. Edgerton, M. Peterson, S. Raza, S. Regmi, N. Papanikolopoulos,
[Online]. Available: https://bsmn.synapse.org/Explore/Data.
C. Weight, The KiTS19 challenge data: 300 kidney tumor cases with clinical
[37] CAMELYON16, 2021, [Online]. Available: https://camelyon16.grand-challenge.
context, CT semantic segmentations, and surgical outcomes, 2019, pp. 1–13.
org/. [Accessed: 11-Dec-2021].
[68] Analysis of images to detect abnormalities in endoscopy accessed, 2021,
[38] D. Racoceanu, F. Capron, Semantic integrative digital pathology: Insights into
[Online]. Available: https://isbi-aida.grand-challenge.org/. [Accessed: 20-Oct-
microsemiological semantics and image analysis scalability, Pathobiology 83
2021].
(2–3) (2016) 148–155.
[69] P. Rajpurkar, J. Irvin, A. Bagul, D. Ding, T. Duan, H. Mehta, B. Yang, K. Zhu,
[39] iciar, 2018-Challenge.grand-challenge, 2021, [Online]. Available: https://
D. Laird, R.L. Ball, C. Langlotz, K. Shpanskaya, M.P. Lungren, A.Y. Ng, MURA:
iciar2018-challenge.grand-challenge.org/Dataset/. [Accessed: 11-Oct-2021].
Large dataset for abnormality detection in musculoskeletal radiographs, 2017.
[40] I.C. Moreira, I. Amaral, I. Domingues, A. Cardoso, M.J. Cardoso, J.S. Cardoso, [70] Lower spine 3D multi-modality MRI data sets, 2021, [Online]. Available: https:
INbreast: Toward a full-field digital mammographic database, Acad. Radiol. 19 //ivdm3seg.weebly.com/data.html. [Accessed: 20-Oct-2021].
(2) (2012) 236–248. [71] S.S. Halabi, L.M. Prevedello, J. Kalpathy-Cramer, A.B. Mamonov, A. Bilbily, M.
[41] CAMELYON17, 2021, [Online]. Available: https://camelyon17.grand-challenge. Cicero, I. Pan, L.A. Pereira, R.T. Sousa, N. Abdala, et al., The RSNA pediatric
org/. [Accessed: 17-Dec-2021]. bone age machine learning challenge, Radiology 290 (2) (2019) 498–503.
[42] J. Irvin, P. Rajpurkar, M. Ko, Y. Yu, S. Ciurea-Ilcus, C. Chute, H. Marklund, [72] N. Bien, P. Rajpurkar, R.L. Ball, J. Irvin, A. Park, E. Jones, M. Bereket,
B. Haghgoo, R. Ball, K. Shpanskaya, J. Seekins, D.A. Mong, S.S. Halabi, J.K. B.N. Patel, K.W. Yeom, K. Shpanskaya, S. Halabi, E. Zucker, G. Fanton, D.F.
Sandberg, R. Jones, D.B. Larson, C.P. Langlotz, B.N. Patel, M.P. Lungren, A.Y. Amanatullah, C.F. Beaulieu, G.M. Riley, R.J. Stewart, F.G. Blankenberg, D.B.
Ng, CheXpert: A large chest radiograph dataset with uncertainty labels and Larson, R.H. Jones, C.P. Langlotz, A.Y. Ng, M.P. Lungren, Deep-learning-assisted
expert comparison, 2019. diagnosis for knee magnetic resonance imaging: Development and retrospective
[43] A.E.W. Johnson, T.J. Pollard, N.R. Greenbaum, M.P. Lungren, C. Deng, Y. Peng, validation of MRNet, PLoS Med. 15 (11) (2018) 1–19.
Z. Lu, R.G. Mark, S.J. Berkowitz, S. Horng, MIMIC-CXR-JPG, a large publicly [73] L. Deleskog, N. Laursen, B.R. Nielsen, P. Schwarz, Vertebral fracture assessment
available database of labeled chest radiographs, 14, (2019) 1–7. by DXA is inferior to X-ray in clinical severe osteoporosis, Osteoporos. Int. 27
[44] P.F. Raudaschl, P. Zaffino, G.C. Sharp, M.F. Spadea, A. Chen, B.M. Dawant, (7) (2016) 2317–2326.
T. Albrecht, T. Gass, C. Langguth, M. Lüthi, et al., Evaluation of segmentation [74] S. Rezaei, A. Emami, N. Karimi, S. Samavi, Gland segmentation in
methods on head and neck CT: auto-segmentation challenge 2015, Med. Phys. histopathological images by deep neural network, 1–5.
44 (5) (2017) 2020–2036. [75] N. Gessert, M. Nielsen, M. Shaikh, R. Werner, A. Schlaefer, Skin lesion
[45] Pet-radiomics-challenges, 2021, [Online]. Available: https://www.kaggle.com/ classification using ensembles of multi-resolution EfficientNets with meta data,
c/pet-radiomics-challenges/data. [Accessed: 28-Dec-2021]. MethodsX 7 (2020) 100864.
[46] Data-science-bowl-2017, 2021, [Online]. Available: https://www.kaggle.com/c/ [76] Skin cancer MNIST, 2021, [Online]. Available: https://www.kaggle.com/
data-science-bowl-2017. [Accessed: 14-Jul-2021]. kmader/skin-cancer-mnist-ham10000. [Accessed: 04-Nov-2021].
[47] S.G. Armato III, G. McLennan, L. Bidaut, M.F. McNitt-Gray, C.R. Meyer, A.P. [77] P. Bilic, P.F. Christ, E. Vorontsov, G. Chlebus, H. Chen, Q. Dou, C.-W. Fu, X.
Reeves, B. Zhao, D.R. Aberle, C.I. Henschke, E.A. Hoffman, et al., The lung Han, P.-A. Heng, J. Hesser, et al., The liver tumor segmentation benchmark
image database consortium (LIDC) and image database resource initiative (lits), 2019, arXiv Prepr. arXiv:1901.04056.
(IDRI): a completed reference database of lung nodules on CT scans, Med. Phys. [78] X. Liu, L. Yang, J. Chen, S. Yu, K. Li, Region-to-boundary deep learning model
38 (2) (2011) 915–931. with multi-scale feature fusion for medical image segmentation, Biomed. Signal
[48] Lung nodule analysis, 2020, [Online]. Available: https://lumic.grand-challenge. Process. Control 71 (2022) 103165.
org/. [Accessed: 11-May-2020]. [79] X. Yan, H. Tang, S. Sun, H. Ma, D. Kong, X. Xie, AFTer-UNet: Axial fusion
[49] Lung nodule analysis-2016, 2020, [Online]. Available: https://luna16.grand- transformer unet for medical image segmentation, in: Proceedings of the
challenge.org/. [Accessed: 11-May-2020]. IEEE/CVF Winter Conference on Applications of Computer Vision, WACV, 2022,
[50] J. Borovec, A. Munoz-Barrutia, J. Kybic, Benchmarking of image registra- pp. 3971–3981.
tion methods for differently stained histological slides, in: 2018 25th IEEE [80] Z. Zhang, C. Wu, S. Coleman, D. Kerr, DENSE-inception U-net for medical image
International Conference on Image Processing, ICIP, 2018, pp. 3368–3372. segmentation, Comput. Methods Programs Biomed. 192 (2020) 105395.

11
P.K. Mall, P.K. Singh, S. Srivastav et al. Healthcare Analytics 4 (2023) 100216

[81] Y. Al Khalil, S. Amirrajab, C. Lorenz, J. Weese, J. Pluim, M. Breeuwer, On [106] K. Simonyan, A. Zisserman, Very deep convolutional networks for large-scale
the usability of synthetic data for improving the robustness of deep learning- image recognition, 2014, arXiv Prepr. arXiv:1409.1556.
based segmentation of cardiac magnetic resonance images, Med. Image Anal. [107] Y. Wu, Y. Xia, Y. Song, Y. Zhang, W. Cai, NFN+: A novel network followed
84 (2023) 102688. network for retinal vessel segmentation, Neural Netw. 126 (2020) 153–162.
[82] F. Yuan, Z. Zhang, Z. Fang, An effective CNN and transformer complementary [108] H. Riaz, J. Park, H. Choi, H. Kim, J. Kim, Deep and densely connected networks
network for medical image segmentation, Pattern Recognit. 136 (2023) 109228. for classification of diabetic retinopathy, Diagnostics 10 (1) (2020) 24.
[83] Y. Wu, K. Liao, J. Chen, J. Wang, D.Z. Chen, H. Gao, J. Wu, D-former: a U- [109] B.H.M. van der Velden, H.J. Kuijf, K.G.A. Gilhuijs, M.A. Viergever, Explainable
shaped dilated transformer for 3D medical image segmentation, Neural Comput. artificial intelligence (XAI) in deep learning-based medical image analysis, 2021,
Appl. 35 (2) (2023) 1931–1944. arXiv Prepr. arXiv:2107.10912.
[84] L. Bonaldi, A. Pretto, C. Pirri, F. Uccheddu, C.G. Fontanella, C. Stecco, Deep [110] Y.D. Kim, K.J. Noh, S.J. Byun, S. Lee, T. Kim, L. Sunwoo, K.J. Lee, S.-H. Kang,
learning-based medical images segmentation of musculoskeletal anatomical K.H. Park, S.J. Park, Effects of hypertension, diabetes, and smoking on age and
structures: A survey of bottlenecks and strategies, Bioengineering 10 (2) (2023). sex prediction from retinal fundus images, Sci. Rep. 10 (1) (2020) 1–14.
[85] K. Chaitanya, E. Erdil, N. Karani, E. Konukoglu, Local contrastive loss [111] C. Spampinato, S. Palazzo, D. Giordano, M. Aldinucci, R. Leonardi, Deep
with pseudo-label based self-training for semi-supervised medical image learning for automated skeletal bone age assessment in X-ray images, Med.
segmentation, Med. Image Anal. 87 (2023) 102792. Image Anal. 36 (2017) 41–51.
[86] C. Sridhar, P.K. Pareek, R. Kalidoss, S.S. Jamal, P.K. Shukla, S.J. Nuagah, [112] D.-H. Lee, et al., Pseudo-label: The simple and efficient semi-supervised learning
Optimal medical image size reduction model creation using recurrent neural method for deep neural networks, in: Workshop on Challenges in Representation
network and GenPSOWVQ, J. Healthc. Eng. 2022 (2022) 2354866. Learning, ICML, Vol. 3, 2013, no. 2.
[87] H. Wang, S. Xie, L. Lin, Y. Iwamoto, X.-H. Han, Y.-W. Chen, R. Tong, Mixed [113] D.B. Larson, M.C. Chen, M.P. Lungren, S.S. Halabi, N.V. Stence, C.P. Langlotz,
transformer U-net for medical image segmentation, in: ICASSP 2022-2022 IEEE Performance of a deep-learning neural network model in assessing skeletal
International Conference on Acoustics, Speech and Signal Processing, ICASSP, maturity on pediatric hand radiographs, Radiology 287 (1) (2018) 313–322.
2022, pp. 2390–2394. [114] S. Koitka, A. Demircioglu, M.S. Kim, C.M. Friedrich, F. Nensa, Ossification area
[88] S. Ren, K. He, R. Girshick, J. Sun, Faster r-cnn: Towards real-time object
localization in pediatric hand radiographs using deep neural networks for object
detection with region proposal networks, Adv. Neural Inf. Process. Syst. 28
detection, PLoS One 13 (11) (2018) e0207496.
(2015) 91–99.
[115] S.H. Tajmir, H. Lee, R. Shailam, H.I. Gale, J.C. Nguyen, S.J. Westra, R. Lim,
[89] A. Ismail, T. Rahmat, S. Aliman, Chest X-ray image classification using faster
S. Yune, M.S. Gee, S. Do, Artificial intelligence-assisted interpretation of bone
R-CNN, Malays. J. Comput. 4 (1) (2019) 225–236.
age radiographs improves accuracy and decreases variability, Skeletal Radiol.
[90] E.M.A. Anas, P. Mousavi, P. Abolmaesumi, A deep learning approach for real
48 (2) (2019) 275–283.
time prostate segmentation in freehand ultrasound guided biopsy, Med. Image
[116] Y. LeCun, et al., Generalization and network design strategies, Connect.
Anal. 48 (2018) 107–116.
Perspect. 19 (1989) 143–155.
[91] V.D.P. Jasti, A.S. Zamani, K. Arumugam, M. Naved, H. Pallathadka, F. Sammy,
[117] V.I. Iglovikov, A. Rakhlin, A.A. Kalinin, A.A. Shvets, Paediatric bone age
A. Raghuvanshi, K. Kaliyaperumal, Computational technique based on machine
assessment using deep convolutional neural networks, in: Lect. Notes Comput.
learning and image processing for medical image analysis of breast cancer
Sci. (Including Subser. Lect. Notes Artif. Intell. Lect. Notes Bioinformatics), in:
diagnosis, Secur. Commun. Netw. 2022 (2022) 1918379.
LNCS, vol. 11045, 2018, pp. 300–308.
[92] P. Kora, C.P. Ooi, O. Faust, U. Raghavendra, A. Gudigar, W.Y. Chan, K.
[118] J.C. Castillo, Y. Tong, J. Zhao, F. Zhu, RSNA bone-age detection using transfer
Meenakshi, K. Swaraja, P. Plawiak, U. Rajendra Acharya, Transfer learning
learning and attention mapping, 2018.
techniques for medical image analysis: A review, Biocybern. Biomed. Eng. 42
[119] P. Van Molle, M. De Strooper, T. Verbelen, B. Vankeirsbilck, P. Simoens,
(1) (2022) 79–107.
[93] B.D. de Vos, F.F. Berendsen, M.A. Viergever, H. Sokooti, M. Staring, I. Išgum, B. Dhoedt, Visualizing convolutional neural networks to improve decision
A deep learning framework for unsupervised affine and deformable image support for skin lesion classification, in: Understanding and Interpreting Ma-
registration, Med. Image Anal. 52 (2019) 128–143. chine Learning in Medical Image Computing Applications, Springer, 2018, pp.
[94] J. Chen, E.C. Frey, Y. He, W.P. Segars, Y. Li, Y. Du, TransMorph: Transformer 115–123.
for unsupervised medical image registration, Med. Image Anal. 82 (2022) [120] I. Kandel, M. Castelli, A. Popovič, Musculoskeletal images classification for
102615. detection of fractures using transfer learning, J. Imaging 6 (11) (2020) 127.
[95] B.A. Skourt, A. El Hassani, A. Majda, Lung CT image segmentation using deep [121] H. Lee, S. Tajmir, J. Lee, M. Zissen, B.A. Yeshiwas, T.K. Alkasab, G. Choy, S.
neural networks, Procedia Comput. Sci. 127 (2018) 109–113. Do, Fully automated deep learning system for bone age assessment, J. Digit.
[96] F.T. Ferreira, P. Sousa, A. Galdran, M.R. Sousa, A. Campilho, End-to-end Imaging 30 (4) (2017) 427–441.
supervised lung lobe segmentation, in: 2018 International Joint Conference on [122] T. Van Steenkiste, J. Ruyssinck, O. Janssens, B. Vandersmissen, F. Vande-
Neural Networks, IJCNN, 2018, pp. 1–8. casteele, P. Devolder, E. Achten, S. Van Hoecke, D. Deschrijver, T. Dhaene,
[97] Y. Oh, S. Park, J.C. Ye, Deep learning covid-19 features on cxr using limited Automated assessment of bone age using deep learning and Gaussian process
training data sets, IEEE Trans. Med. Imaging 39 (8) (2020) 2688–2700. regression, in: Proc. Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. EMBS, Vol.
[98] F. Ucar, D. Korkmaz, COVIDiagnosis-Net: Deep Bayes-SqueezeNet based diag- 2018-July, 2018, pp. 674–677.
nosis of the coronavirus disease 2019 (COVID-19) from X-ray images, Med. [123] R. Mishra, O. Daescu, P. Leavey, D. Rakheja, A. Sengupta, Histopathological
Hypotheses 140 (2020) 109761. diagnosis for viable and non-viable tumor prediction for osteosarcoma using
[99] X. Wu, H. Hui, M. Niu, L. Li, L. Wang, B. He, X. Yang, L. Li, H. Li, J. convolutional neural network, in: International Symposium on Bioinformatics
Tian, et al., Deep learning-based multi-view fusion model for screening 2019 Research and Applications, 2017, pp. 12–23.
novel coronavirus pneumonia: A multicentre study, Eur. J. Radiol. 128 (2020) [124] S. Mahore, K. Bhole, S. Rathod, Machine learning approach to classify and
109041. predict different osteosarcoma types, in: 2021 8th International Conference on
[100] H. Behzadi-Khormouji, H. Rostami, S. Salehi, T. Derakhshande-Rishehri, M. Signal Processing and Integrated Networks, SPIN, 2021, pp. 641–645.
Masoumi, S. Salemi, A. Keshavarz, A. Gholamrezanezhad, M. Assadi, A. Batouli, [125] B.H.M. van der Velden, H.J. Kuijf, K.G.A. Gilhuijs, M.A. Viergever, Explainable
Deep. learning, Reusable and problem-based architectures for detection of artificial intelligence (XAI) in deep learning-based medical image analysis, Med.
consolidation on chest X-ray images, Comput. Methods Programs Biomed. 185 Image Anal. 79 (2022) 102470.
(2020) 105162. [126] B. Davazdahemami, H.M. Zolbanin, D. Delen, An explanatory analytics frame-
[101] A. Mittal, D. Kumar, M. Mittal, T. Saba, I. Abunadi, A. Rehman, S. Roy, work for early detection of chronic risk factors in pandemics, Healthc. Anal. 2
Detecting pneumonia using convolutions and dynamic capsule routing for chest (2022) 100020.
X-ray images, Sensors 20 (4) (2020) 1068. [127] H.B. Arunachalam, R. Mishra, O. Daescu, K. Cederberg, D. Rakheja, A. Sengupta,
[102] S. Sabour, N. Frosst, G.E. Hinton, Dynamic routing between capsules, 2017, D. Leonard, R. Hallac, P. Leavey, Viable and necrotic tumor assessment from
arXiv Prepr. arXiv:1710.09829. whole slide images of osteosarcoma using machine-learning and deep-learning
[103] R. Boddeda, S.V. Deepak, S.A. Patel, et al., A novel strategy for COVID-19 models, PLoS One 14 (4) (2019) 1–19.
classification from chest X-ray images using deep stacked-ensembles, 2020, [128] M. D’Acunto, M. Martinelli, D. Moroni, Deep learning approach to human
arXiv Prepr. arXiv:2010.05690. osteosarcoma cell detection and classification, in: International Conference on
[104] R. Wang, S. Chen, C. Ji, J. Fan, Y. Li, Boundary-aware context neural network Multimedia and Network Information System, 2018, pp. 353–361.
for medical image segmentation, Med. Image Anal. 78 (2022) 102395. [129] R. Sayres, A. Taly, E. Rahimy, K. Blumer, D. Coz, N. Hammel, J. Krause, A.
[105] D. Nagasato, H. Tabuchi, H. Masumoto, H. Enno, N. Ishitobi, M. Kameoka, Narayanaswamy, Z. Rastegar, D. Wu, et al., Using a deep learning algorithm
M. Niki, Y. Mitamura, Automated detection of a nonperfusion area caused by and integrated gradients explanation to assist grading for diabetic retinopathy,
retinal vein occlusion in optical coherence tomography angiography images Ophthalmology 126 (4) (2019) 552–564.
using deep learning, PLoS One 14 (11) (2019) e0223965.

12