Evaluation of Machine Learning Architectures in

Healthcare
Yash Verma Shahab Tayeb
Department of Electrical and Computer Engineering Department of Electrical and Computer Engineering
California State University California State University
2021 IEEE 11th Annual Computing and Communication Workshop and Conference (CCWC) | 978-1-6654-1490-6/21/$31.00 ©2021 IEEE | DOI: 10.1109/CCWC51732.2021.9375834

Fresno, USA Fresno, USA

[email protected] [email protected]

Abstract—Machine Learning (ML) is now influencing every The present survey is based on nineteen papers based on
part of the industry. From detecting objects from an image to different proposed CNN models to detect various diseases. All
recommending different items while doing online shopping based the conditions are fatal and common. These papers were
on someone's recent browsing history. Now, the ML is touching published within the last five years. Each article has tried to
the healthcare sector. It is now an area of interest for more doctors implement a different ML version or introduced a new method
and scientists to implement different techniques and harvest the to do so, or customized the previous algorithms.
power of ML. Over the past few years, there is a race to implement
Artificial Intelligence (AI) and ML in this sector. Multiple scholars In this survey, different papers have been put together into
have presented their approach. Currently, the proposed models diverse groups of likeness to provide a rational comparison of
are nowhere near the actual implementation of these models in the the proposed methodologies. Some of the papers and the
real world. However, these models are laying down the path to do methods discussed in them have some similarities, making it
so in the future. Here is a review of some of the papers discussing interesting to contrast how one approach works with one
different techniques for targeting various diseases using AI/ML. situation and how it works with a different one.
Each paper introduces a method developed based on the separate
datasets created from organically collected data. II. DATASET
Keywords—Machine Learning, Artificial Intelligence, Datasets are an essential commodity for training a model to
Pneumonia, transfer learning, automation do something new. The data fed into the model to train upon
decides the outcome quality of the model. Better the data's
I. INTRODUCTION quality with annotation of the focused area, better will be the
accuracy, precision, and sensitivity. The various models
Since the inception of machine learning and artificial
discussed here are trained upon several different datasets. Some
intelligence, it has automated different processes and limited
of these datasets are privately owned and managed, and some
human intervention. ML is getting very smart over time, as the
are publicly available for future improvements. All the
data to learn from is also increasing. In all modern hospitals,
information contained in these datasets has been generalized to
patient data is kept for future reference. As this data increases,
maintain the privacy of the subjects. Most of the pre-trained
the chances of implementing an automated system to detect,
models discussed here are trained upon a master dataset called
classify, and segment the abnormality is also increasing. Some
ImageNet [2]. ImageNet is a widely known dataset used to train
of the data scientists and doctors have already started working
the CNN models. It has more than 14 million images with 1
on such systems with the available data [1].
million images with bounding boxes that help models during
The motivation behind this study is to compare which training. Then these pre-trained models are later modified for the
architecture works best and has the highest accuracy. The model specific domain. Later on, these modified pre-trained models are
will take an image as an input and process the input, and tell again trained on the specific datasets for that particular domain
whether the image has an anomaly. If succeeded in implementation. These datasets sizes are essential because
implementing such systems for the public, it will fast forward larger datasets provide an accurate mean value, helps in
the whole process until the anomaly detection. It will be cheaper, identifying outliers that could modify the accuracy and
accurate, and safe as there will be less human error. It will also precision, which will, in the long run, help in reducing the error
reduce the physician's burnout caused by an exhaustive schedule margin. The following table gives a brief description of the
to see, diagnose, and treat patients. These systems can also be public datasets used by the models proposed by the different
accessed from remote locations, which will help in providing authors.
findings instantly.
TABLE I. SOME OF THE DATASETS USED FOR TRAINING DIFFERENT
Multiple papers have been discussed here about different MODELS
abnormalities and how ML and AI's power has been used to
detect, classify, and segment those abnormalities. These Dataset Dataset
Dataset name Date created
abnormalities can later be detected as a particular disease with owner summary
an advanced diagnosis occurring in various body parts. These Chest X-Ray Paul 03/22/2018 Total images:
papers focus more on the cranial cavity and the thoracic cavity. Images [3] Mooney 5,863 (624 test
images, 5,216

978-0-7381-4394-1/21/$31.00 ©2021 IEEE

1377

Authorized licensed use limited to: UNIVERSITY OF NEW MEXICO. Downloaded on May 16,2021 at 14:11:50 UTC from IEEE Xplore. Restrictions apply.
 train images, 16 whole system has 13 layers, out of which 9 are convolutional
images for layers, 3 are fully connected layers, and a classification layer.
validity) The system and implementation are sound, and the details of the
whole system's dataset are nicely presented. The details of how
COVID-19 T. 03/29/2020 Total images: the dataset was prepared, how the pre-processing was done, and
Radiography Rahman 3,886 (1200 the dataset source are all missing. All these questions make the
Database [4] et al. COVID-19 credibility of the dataset uncertain.
positive images,
1341 normal Lastly, Cho et al. [11] introduced the system to automate
images, and 1345 Hydrops ration calculation in Meniere's disease. The primary
viral pneumonia purpose of the system was to develop to do the segmentation.
images) Modality is MRI. The main highlight of this system is that they
have applied a gradual learning process. So the system first
NIH ChestX- National 12/02/2017 Total images: learns from easy problems and then moves towards the harder
Ray14 [5] Institute 112,120. ones. The best performing number was achieved by the system
of Health Compromise 15 in experiment 5 with 372 patches, having an intersection-over-
Chest X- classes (14 for union value of 0.761±0.036. The results were good, but the
Ray diseases and 1 for dataset on which the system was trained was comparably small.
dataset normal) Due to the smaller size of the dataset, the trained system cannot
CheXpert: A J. Irvin et 01/21/2019 Total images: be efficient for a large population with countless variations.
large chest x- al. 224,316 B. Thoracic Cavity Group
ray dataset
and The thoracic cavity is the part up from the abdominal cavity
competition where the lungs and heart are present. Most of the papers in this
[6] survey are based on lung diseases like Pneumonia and Lung
cancer, hence lies in this category.
Labhane et al. [12] proposed a system to detect pediatric
III. TAXONOMY pneumonia from the lung's X-ray images. In his system,
Labhane used transfer learning architecture like VGG16,
A. Cranial Cavity group VGG19, InceptionV3, under which the knowledge gained from
The cranial cavity is the upper part of the body, including the solving one problem is used while solving another problem. This
head, brain, nose, eyes, ears, and buccal cavity. Three papers in is a recent study published in 2020 CE only. The best results
the survey lie in this category. were observed with InceptionV3. It has an accuracy of 98%,
recall value of 0.98, F1-score of 0.98, and precision of 0.99. The
Seetha et al. [7] discussed the classification of brain tumors proposed method produced desirable results at 100 epochs. The
using MRI as a modality. The paper details how the neural dataset on which the model is trained is Mendeley Data2 [13].
network functions and how a convolutional neural network can
help classify the disease. The proposed method does not contain Chowdhury et al. [14] developed a model to detect a typical
any predefined CNN modules, but the paper presents a new viral or COVID-19 Pneumonia. This study is also a very recent
algorithm to implement the system. The system has a training study based on the situation that occurred during the year 2020
accuracy of 97.5%. Moreover, the datasets used are publicly pandemic. It was getting harder to differentiate between the two
available, called the Radiopaedia dataset and Brain Tumor diseases as both of them have almost similar symptoms. Taking
Image Segmentation Benchmark (BRATS) 2015 testing dataset care of this in mind, Muhammed and his team developed this
[8][9]. The system is comparably faster because of its system. X-ray images were used as a modality. This study
lightweight, but nowhere in the paper is that they have compared different pre-trained Neural Networks (NN), which
considered training the system on a healthy brain; that is why were made to train on the datasets gathered from various private
the proposed system's precision is questionable. and public sources. The proposed method has two variations,
one with 121 layers of DenseNet and another with 201 layers of
Dolz et al. [10] proposed three systems to do the DenseNet. DenseNet is a model that is used for visual object
segmentation on an infant's brain. The proposed system again recognition. The whole system has 20 backpropagation epochs.
uses an MRI to train itself. Here the system is based on a 3D The results for normal, COVID-19, and viral pneumonia are:
Fully Convolutional Neural Network (FCNN) architecture. accuracy: 97.9%, precision: 97.95%, sensitivity: 97.9%, and
There are three methods proposed, EarlyFusion_Single, specificity: 98.8%.
EarlyFusion_Ensemble, and LateFusion_Ensemble. Three
metrics were used to evaluate the proposed systems' accuracy, Saric et al. [15] proposed a system where they used
i.e., Dice Similarity Coefficient, Modified Hausdorff Distance, Histopathology images. The proposed approach was based on
and Average Surface Distance. All the systems were in the top CNN based method to detect and classify Lung Cancer. The
five best-performing systems where these systems were VGG16 architecture achieved the highest patch classification
presented. The proposed system is faster in segmenting the white accuracy of 0.7541. The main focus of this system is that it uses
matter, grey matter, and cerebrospinal fluid. The FCNN two CNN architectures to classify one patch. This
implemented takes 30 epochs, each with 20 sub-epochs. The implementation gives conformity to the system's output, but this

1378

Authorized licensed use limited to: UNIVERSITY OF NEW MEXICO. Downloaded on May 16,2021 at 14:11:50 UTC from IEEE Xplore. Restrictions apply.
double processing puts a burden on the system and increases the i.e., 0.962. The paper goes much more in detail while elaborating
complexity. Saric has used pre-trained deep CNNs like VGG16 on different tasks, models, data, and data visualization.
and ResNet50. The whole system has 16 layers and 50 layers
variant, but both variants have 17 epochs. The dataset used here Gu et al. [24] took the pneumonia detection study to another
is from ACDC@LUNGHP [16], and multiple patches are level. They developed a system that can classify bacterial and
created from 25 images. viral childhood pneumonia using X-Ray images as a modality.
The accuracy is 0.8048 ± 0.0202 and sensitivity is 0.7755 ±
Varshni et al. [17] proposed and developed a system to detect 0.0296. However, their dataset on which they have trained their
pneumonia using CNN feature extraction. The modality here is AlexNet based Fully Convolutional Network is comparably
X-ray images. Here they have used a pre-trained system called small, making the model's output on a bigger subject group
DenseNet-169, which has densely connected 169 layers. The questionable. It also compares different methods like GLCM
paper gives useful details on how they have worked with Features, Wavelet Features, HOG Features, and others. The
DenseNet-169. However, for this system to execute, it needs dataset contains only 4892 x-ray images. The proposed model
high computational power. The highest AUC achieved by the has eight convolutional layers and three fully connected layers,
feature extractor, DenseNet-169, with classifier SVM (rbf and the model takes 120 epochs to process each image.
kernel) with gamma 1.9e-05 is 0.8002. The dataset used by them
Togacar et al. [25] developed a pneumonia detection model
is publicly available called ChestX-ray14 [5].
applying the combination of mRMR feature selection. mRMR
Sirazitdinov et al. [18] proposed a system to detect and is frequently used for feature recognition in gnomes and
classify the disease pneumonia using X-rays as a modality. They phenotype. In their model, 1000 features were selected using
have used a Deep Neural Network (DNN) and focuses on large different classifiers. Since the model has multiple classifiers,
scale database handling and processing. They have used computation time is high and requires powerful hardware to
RetinaNet and Mask R-CNN. Moreover, the dataset used for the handle the processing. The best results were obtained in terms
training is the RSNA Pneumonia Detection Challenge dataset, of accuracy was obtained using AlexNet & VGG-16 & VGG-19
as the solution was introduced in a competition where the dataset with LDA classifier, with a high of 99.41 with an augmented
was provided to them. A comparison of the proposed dataset. The model has different CNN models used parallel to
architecture with DenseNet-121 and ResNet-50 is also given in each other, with varying counts of the layer. The model was
the paper. trained on a public ChestX-ray dataset of the volunteer patients.
Zech et al. [19] publish another paper to detect and classify Kermany et al. [26] also presented their study in which they
Pneumonia from X-rays. Their study collected the data from showed their general view on diagnosis based on image-based
three credible sources and put their model to train on them. Zech deep learning. It lays a foundation for future studies in the
and his team have done a fabulous job of putting and comparing healthcare field. The accuracy attained by the model they
their model results on these three datasets. The best result is developed is 92.8%. Sensitivity is 93.2%, and the specificity of
observed when they have used superset in comparison type in 90.1%. The learning model is based on the InceptionV3, a pre-
MSH + NIH train site, i.e., 0.732 (0.727 – 0.737). trained learning model, and they have also applied transfer
learning on top of it. The model takes 100 epochs for the
Xu et al. [20] developed a system called CXNet-m1 for processing.
detecting anomalies in Chest X-rays by harvesting the power of
Image-based deep learning. Initially, their model's results were Al Mamlook et al. [27] studied pneumonia and questioned
not impressive as they were all below 75%. They developed each aspect of the detection process and, in the end, came up
three different versions of their own systems CXNet-m1 (v0), with their model based on Keras with Tensorflow. The model
which is also the original model, CXNet-m1 (v1), which also has eight convolutional layers. The accuracy of their model is
includes sin-loss, and CXNet-m1 (v2), which is a new model 98.46%. The model takes 100 epochs for each image processing
plus sin-loss. Now, after splitting their models' highest accuracy and uses a Decision tree, KNN, and Random Forest for the
achieved is 93.6% on OpenI [21]. validation. Being the model having a good accuracy, it is trained
on a limited dataset. It cannot be implemented for general public
Cohen et al. [22] thought of providing a web-based use without training it more on a bigger dataset.
application to do data processing locally on the system. From
the point of privacy, it is a good idea, as the processing is done Wang et al. [28] developed a hospital-scale supervised multi-
locally. Thus, the system will need a good and constant internet label classification model. This study tries to widen the range to
connection that might not be available in remote locations. The all common thorax diseases, and for that, they have used chest
highest AUC (performance) of 0.93 ± 0.01 is observed for the x-ray images as a modality. They have used multiple pre-trained
disease Emphysema, and the lowest is 0.72 ± 0.01 for CNNs like AlexNet, GoogLeNet, VGGNet-16, and ResNet-50
pneumonia. and gives a fair assessment of using different CNNs concerning
the different diseases. The best performance was achieved with
Rajaraman et al. [23] developed a system specifically to ResNet-50 with all the diseases except one. The highest
visualize the data using a convolutional neural network. They accuracy was achieved for Cardiomegaly disease at 0.9931 at
used a customized VGG16 CNN model. It was good to visualize T(Intersection Over the detected B-Box area ratio) equals 0.1.
data to understand it more nicely, but the model was They have used ChestX-ray8 and OpenI [21] dataset. These are
computationally complex. Due to this, a sound system would be big datasets, so one needs to implement an efficient data
required to run the model. The highest accuracy was achieved cleansing and pre-processing system.
between the task of comparing Normal v. Pneumonia images,

1379

Authorized licensed use limited to: UNIVERSITY OF NEW MEXICO. Downloaded on May 16,2021 at 14:11:50 UTC from IEEE Xplore. Restrictions apply.
 Sa et al. [29] developed a system to detect abnormalities in Saric et al. VGG16, 75.41%,
- -
the spine's disc placement. It is a fascinating study, but the [15] ResNet50 72.05%
RetinaNet +
dataset of spinal x-ray images on which the proposed model is Mask
trained is small. The accuracy of the model is always above 0.98. 79.3%, 75.8%,
Sirazitdinov RCNN,
- 65.2%, 88.3%,
et al. [18] DenseNet-
Mo et al. [30] proposed a model to classify multi-labeled 56.9% 85.5%
121,
images using entropy weight loss to observe inter-label ResNet-50
independencies. A pre-trained model called DenseNet-121 was 73.2% 95%
ResNet-50,
Zech et al. (72.7% – (94.2% –
used, having 121 layers. The model works in 50 epochs. The [19]
DenseNet-
73.7%) 95.7%)
-
ChestX-ray14 dataset was used to train the model. The average 121
(highest) (highest)
AUC value for the model is 0.8430. Some tweaks were made in CXNet-
the model algorithms to detect other abnormalities, but a m1(v0),
93.1%, 89.9%, 93.9%,
significant improvement was observed for 'Hernia.' Xu et al. [20] CXNet-
93.6% 93.0% 92.4%
m1(v1)
Akpinar et al. [31] recently proposed a chest abnormality (For OpenI)
Customized
detection model that uses SqueezeNet. SqueezeNet is a deep VGG16
neural network for computer vision and takes x-ray images as an Rajaraman et Kermany et
96.2% 96.2% 97.7%
input. The method also focuses on the pre-processing of the al.[23] al.
images. The dataset used is ChestX-ray14. The model's accuracy (Normal v.
Pneumonia)
is 90.95%, but the paper does not mention its specificity and AlexNet,
sensitivity. based FCN 80.48±0.02 77.55±0.02 74.06±0.04
Gu et al. [24]
(DCNN 02% 96% 52%
IV. FINDINGS Features)
AlexNet +
From the models observed and mentioned in the last section, VGG-16 +
we can easily conclude a few similarities between different Togacar et
VGG-19, 99.41% 99.61% 99.22%
al. [25]
models. A thorough look at the literature reveals that VGG-16, with LDA
AlexNet, ResNet, and InceptionV3 are among the highly-used classifier
Kermany et
pre-trained models used for detection and classification al.[26]
InceptionV3 92.8% 93.2% -
purposes. Al Mamlook Keras with
98.46% - -
et al. [27] Tensorflow
ChestX-ray14 [5] dataset is among the most widely used 79% (for 100% (for
dataset for the Thoracic cavity diagnosis. ChestX-ray14 contains 99.31% Cardiomega Cardiomega
over 112,000 chest x-ray images from more than 30,000 unique AlexNet,
(for ly), ly),
patients. National Institute of Health created it. Cardiomeg 93% (for 90% (for
GoogLeNet,
aly) and pneumonia) pneumonia)
Wang et VGG16,
75% (for (evaluation (evaluation
al.[28] ResNet-50
Pneumonia of image of image
(highest
) at labelling labelling
value)
T(IoBB)=0. results on results on
1 OpenI OpenI
dataset) dataset)
90.5%
(highest by
Faster-
Sa et al.[29] Above 98% - Faster-
RCNN
RCNN
(large)
Akpinar et
SqueezeNet 90.95% - -
al. [31]

From the above table, we can see a pattern in the uses of the
different architectures. VGG16, DenseNet, and InceptionV3 are
Fig. ChesX-ray14 dataset sample images.
among the most used CNN architecture here. Let us discuss
The primary modality is also the X-ray images for many these more.
papers and proposed models. It can be because of the wide
Labhane et al., Saric et al., Rajaram et al., and Wang et al.
availability of x-ray machines and the low price for developing
used VGG16 in their proposed models [12]. VGG16 model has
these images.
16 layers. VGG16 also has a variation with 19 layers, called
TABLE II. COMPARISON OF PAPERS ON THORACIC CAVITY
VGG19. VGG16 can achieve up to 92.7% accuracy during
training/testing on the ImageNet dataset. It has a small kernel
Author Model Accuracy Sensitivity Precision size, which makes the training on VGG16 a little slow. To run
VGG16, the model, one needs a robust system as it requires high
Labhane et VGG19, 98%, 97%, 96%, 95%, 99%, 99%,
al. [12] and and 98% and 98% and 99% bandwidth and a faster storage system. Labhane et al. developed
InceptionV3 multiple models using Basic CNN, VGG16, VGG19, and
Chowdhury DenseNet-
97.9% 97.9% 97.95% InceptionV3. For VGG16, their accuracy was 98%, with a
et al. [14] 121 sensitivity of 0.98 and precision of 0.99, which is very good.

1380

Authorized licensed use limited to: UNIVERSITY OF NEW MEXICO. Downloaded on May 16,2021 at 14:11:50 UTC from IEEE Xplore. Restrictions apply.
They had a total of 1703 testing images for their model. Saric V. CONCLUSION
also used VGG16 architecture for their proposed model with an The current study gives a brief description of how ML can
accuracy of 75.41%. Rajaraman et al. were able to achieve an help us detect, classify and segment different diseases from
accuracy of 96.2%. Their model's great thing was that they some test result images. It also provides a quick comparison
customized the VGG16 to achieve high accuracy and distinguish between multiple CNN architectures, their model and process,
between viral and bacterial pneumonia, which can further help about their datasets on which they were trained, accuracy,
the clinicians treat and target the disease better. precision, and sensitivity they achieved. This study concludes
Labhane et al. and Kermany et al. developed their systems that VGG, DenseNet, and InceptionV3 architecture are among
based on InceptionV3. InceptionV3 is an extensively used the top contenders for reliable model implementation. As we
image-recognition system that has an accuracy generally higher have seen in different studies, even these architectures can give
than 78.1%. It is lightweight, each to implement, and can varying results based on the dataset they are trained on or due to
achieve the lowest error rate. Labhane, in his model, was able to the simple architecture modification. Table II shows different
get an accuracy of 98%, with a sensitivity of 0.98 and precision results observed for different models. It also needs to be
of 0.99, which is among the highest, as he mentioned in his considered that more the models and architectures are complex
paper. Kermany et al. developed a multi-classifying model. or have more convolutional layers, more powerful systems will
Specifically for pneumonia detection, their accuracy rate was be required for them to function at full efficiency. And, it is also
90.7%, with a sensitivity of 88.6% and a specificity of 90%. vital that these models get trained on big datasets. Still, due to
the limited availability and small sizes of the datasets, we cannot
DenseNet is a model of layers densely connected, and they be sure about these models' accuracy. Though the present
have shorter connections between them, which results in better research gives highly noticeable results, we are still far from
efficiency. It's a forward feed system. Chowdhury et al., Varshni perfecting and implementing these systems in the real world.
et al., Zech et al., and Mo et al. proposed their designs based on More research needs to be done to use them in real-time.
DenseNet. Chowdhury et al. developed multiple models based
on different architectures like SqueezeNet, DenseNet, REFERENCES
InceptionV3, etc. And, interestingly they were able to test these [1] S. Tayeb et al., "Toward predicting medical conditions using k-nearest
model on augmented and non-augmented images. In this neighbors," 2017 IEEE International Conference on Big Data (Big Data),
process, DenseNet-201 achieved the highest accuracy of Boston, MA, 2017, pp. 3897-3903. doi: 10.1109/BigData.2017.8258395
99.70% with image augmentation in a two-class classification; [2] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li and L. Fei-Fei, ImageNet: A
and for a three-class classification problem, the highest accuracy Large-Scale Hierarchical Image Database. IEEE Computer Vision and
is 97.94%, again achieved by DenseNet201. The highest Pattern Recognition (CVPR), 2009
accuracy of 99.41% was achieved without image augmentation [3] Kermany, Daniel; Zhang, Kang; Goldbaum, Michael (2018), “Labeled
Optical Coherence Tomography (OCT) and Chest X-Ray Images for
by CheXNet & ResNet18 in a two-class classification problem; Classification,” Mendeley Data, v2. doi: 10.17632/rscbjbr9sj.2
for three-class classification problems, it is 97.74% achieved by [4] T. R. Muhammad E. H. Chowdhury, A. Khandakar, R. Mazhar, M. A.
CheXNet. CheXNet [32] is a 121-layered CNN explicitly Kadir, Z. B. Mahbub, K. R. Islam, M. S. Khan, A. Iqbal, N. Al-Emadi,
trained on the ChestX-ray14 dataset. Overall, DenseNet was and M. Bin I. Reaz. (2020). COVID-19 Chest X-Ray Database. Accessed
able to grab the highest values. Varshni et al. did excellent work on: Jan. 19, 2021. [Online].
on creating different sets of feature extractors and classifiers and https://www.kaggle.com/tawsifurrahman/covid19-radiography-database
comparing them. The highest AUC, i.e., 0.7749, was achieved [5] Wang, Xiaosong, Peng, Yifan, Lu, Le, Lu, Zhiyong, Bagheri,
Mohammadhadi, and Summers, Ronald M. Chestx-ray8: Hospital-scale
by a ResNet-50 model with a Support Vector Machine (SVM). chest x-ray database and benchmarks on weakly-supervised classification
The second highest value of AUC, i.e., 0.7577, was reached by and localization of common thorax diseases. arXiv preprint
DenseNet-121 with an SVM classifier. Mo et al. developed a arXiv:1705.02315, 2017. Accessed on: Jan. 16, 2021. [Online].
multi-labeled classifier based on DenseNet121 using their https://nihcc.app.box.com/v/ChestXray-NIHCC
proposed loss, which has an average accuracy of 0.8430. If [6] J. Irvin, P. Rajpurkar, M. Ko, Y. Yu, S. Ciurea-Ilcus, C. Chute, H.
taking pneumonia specifically, it has an accuracy of 0.7597. Marklund, B. Haghgoo, R. Ball, K. Shpanskaya, et al., “CheXpert: A large
chest radiograph dataset with uncertainty labels and expert comparison,”
These results were based and compared on different losses. This arXiv preprint arXiv:1901.07031, 2019
paper also compares its results with other models like the
[7] Seetha, J. and S. Raja. "Brain Tumor Classification Using Convolutional
ChestNet [33] model, which has two branches, i.e., classification Neural Networks." Biomedical and Pharmacology Journal 11 (2018):
branch, which is based on the ResNet-152 model, and an 1457-1461
attention branch which is for abnormality detection and has six [8] Menze et al., The Multimodal Brain TumorImage Segmentation
convolutional layers and have the result of 0.7810 (average) for Benchmark (BRATS), IEEE Trans. Med. Imaging, 2015
the customized loss; CNNs + Patch slicing [34] model in which [9] Kistler et. al, The virtual skeleton database: an open access repository for
initially images are fed into a CNN (ResNet) which gives biomedical research and collaboration. JMIR, 2013
abstracted feature maps as the output and then this output is [10] Jose Dolz, Christian Desrosiers, Li Wang, Jing Yuan, Dinggang Shen,
passed through patch slicing layer which divides the input in P Ismail Ben Ayed, "Deep CNN ensembles and suggestive annotations for
infant brain MRI segmentation", Computerized Medical Imaging and
× P patch grid and later the probability for the disease type is Graphics, Volume 79, 2020, 101660, ISSN 0895-6111. doi:
calculated and gives the result of 0.8021 (average) for the 10.1016/j.compmedimag.2019.101660.
customized loss. [11] Cho, Y.S., Cho, K., Park, C.J. et al. Automated measurement of hydrops
ratio from MRI in patients with Ménière's disease using CNN-based
segmentation. Sci Rep 10, 7003 (2020). doi: 10.1038/s41598-020-63887-
8

1381

Authorized licensed use limited to: UNIVERSITY OF NEW MEXICO. Downloaded on May 16,2021 at 14:11:50 UTC from IEEE Xplore. Restrictions apply.
[12] G. Labhane, R. Pansare, S. Maheshwari, R. Tiwari and A. Shukla, Learning in Chest Radiography. In <i>Proceedings of the 3rd
"Detection of Pediatric Pneumonia from Chest X-Ray Images using CNN International Conference on Multimedia and Image Processing</i>
and Transfer Learning," 2020 3rd International Conference on Emerging (<i>ICMIP 2018</i>). Association for Computing Machinery, New
Technologies in Computer Engineering: Machine Learning and Internet York, NY, USA, 88–93. doi: 10.1145/3195588.3195597.
of Things (ICETCE), Jaipur, India, 2020, pp. 85-92. doi: [25] M. Toğaçar, B. Ergen, Z. Cömert, F. Özyurt, "A Deep Feature Learning
10.1109/ICETCE48199.2020.9091755 Model for Pneumonia Detection Applying a Combination of mRMR
[13] Kermany, D. K., and M. Goldbaum. "Labeled optical coherence Feature Selection and Machine Learning Models," IRBM, Volume 41,
tomography (OCT) and Chest X-Ray images for classification." Mendeley Issue 4, 2020, Pages 212-222, ISSN 1959-0318. doi:
Data 2 (2018) 10.1016/j.irbm.2019.19.0061.
[14] M. E. H. Chowdhury et al., "Can AI Help in Screening Viral and COVID- [26] Kermany, Daniel S., et al. "Identifying medical diagnoses and treatable
19 Pneumonia?," in IEEE Access, vol. 8, pp. 132665-132676, 2020. doi: diseases by image-based deep learning." Cell 172.5 (2018): 1122-1131.
10.1109/ACCESS.2020.3010287. [27] Almamlook, Rabia & Chen, Shengfeng & Bzizi, Hanin. (2020).
[15] M. Šarić, M. Russo, M. Stella and M. Sikora, "CNN-based Method for Investigation of the performance of Machine Learning Classifiers for
Lung Cancer Detection in Whole Slide Histopathology Images," 2019 4th Pneumonia Detection in Chest X-ray Images.
International Conference on Smart and Sustainable Technologies 10.1109/EIT48999.2020.9208232.
(SpliTech), Split, Croatia, 2019, pp. 1-4. doi: [28] X. Wang, Y. Peng, L. Lu, Z. Lu, M. Bagheri and R. M. Summers,
10.23919/SpliTech.2019.8783041 "ChestX-Ray8: Hospital-Scale Chest X-Ray Database and Benchmarks
[16] Automatic Cancer Detection and Classification in Whole-slide Lung on Weakly-Supervised Classification and Localization of Common
Histopathology (ACDC@LUNGHP). Accessed on: Jan. 16, 2021. Thorax Diseases," 2017 IEEE Conference on Computer Vision and
[Online]. https://acdc-lunghp.grand-challenge.org/Challenge/ Pattern Recognition (CVPR), Honolulu, HI, 2017, pp. 3462-3471. doi:
[17] D. Varshni, K. Thakral, L. Agarwal, R. Nijhawan and A. Mittal, 10.1109/CVPR.2017.369.
"Pneumonia Detection Using CNN based Feature Extraction," 2019 IEEE [29] R. Sa et al., "Intervertebral disc detection in X-ray images using faster R-
International Conference on Electrical, Computer and Communication CNN," 2017 39th Annual International Conference of the IEEE
Technologies (ICECCT), Coimbatore, India, 2019, pp. 1-7. doi: Engineering in Medicine and Biology Society (EMBC), Seogwipo, 2017,
10.1109/ICECCT.2019.8869364. pp. 564-567. doi: 10.1109/EMBC.2017.8036887.
[18] Ilyas Sirazitdinov, Maksym Kholiavchenko, Tamerlan Mustafaev, Yuan [30] S. Mo and M. Cai, "Deep Learning Based Multi-Label Chest X-Ray
Yixuan, Ramil Kuleev, Bulat Ibragimov, "Deep neural network ensemble Classification with Entropy Weighting Loss," 2019 12th International
for pneumonia localization from a large-scale chest x-ray database," 2019 Symposium on Computational Intelligence and Design (ISCID),
Computers & Electrical Engineering, Volume 78, Pages 388-399, ISSN Hangzhou, China, 2019, pp. 124-127. doi: 10.1109/ISCID.2019.10111.
0045-7906. doi: 10.1016/j.compeleceng.2019.08.004. [31] K. N. Akpinar, S. Genc and S. Karagol, "Chest X-Ray Abnormality
[19] Zech, John R., et al. "Variable generalization performance of a deep Detection Based on SqueezeNet," 2020 International Conference on
learning model to detect pneumonia in chest radiographs: a cross- Electrical, Communication, and Computer Engineering (ICECCE),
sectional study." PLoS medicine 15.11 (2018): e1002683. Istanbul, Turkey, 2020, pp. 1-5. doi:
[20] S. Xu, H. Wu and R. Bie, "CXNet-m1: Anomaly Detection on Chest X- 10.1109/ICECCE49384.2020.9179404
Rays With Image-Based Deep Learning," in IEEE Access, vol. 7, pp. [32] Rajpurkar, P., Irvin, J., Zhu, K., Yang, B., Mehta, H., Duan, T., Ding, D.,
4466-4477, 2019. doi: 10.1109/ACCESS.2018.2885997 Bagul, A., Langlotz, C., Shpanskaya, K. and Lungren, M.P., 2017.
[21] OpenI, Open Access Biomedical Image Search Engine. Accessed on: Jan. CheXnet: Radiologist-level pneumonia detection on chest x-rays with
16, 2021. [Online]. https://openi.nlm.nih.gov/ deep learning. arXiv preprint arXiv:1711.05225.
[22] Joseph Paul Cohen, Paul Bertin, Vincent Frappier, "Chester: A web [33] Wang, Hongyu, and Yong Xia. "Chestnet: A deep neural network for
delivered locally computed chest x-ray Disease Prediction System," classification of thoracic diseases on chest radiography." arXiv preprint
arXiv:1901.11210. arXiv:1807.03058 (2018).
[23] Rajaraman, Sivaramakrishnan, et al. "Visualization and interpretation of [34] Z. Li et al., "Thoracic Disease Identification and Localization with
convolutional neural network predictions in detecting pneumonia in Limited Supervision," 2018 IEEE/CVF Conference on Computer Vision
pediatric chest radiographs." Applied Sciences 8.10 (2018): 1715. and Pattern Recognition, Salt Lake City, UT, 2018, pp. 8290-8299. doi:
10.1109/CVPR.2018.00865
[24] Xianghong Gu, Liyan Pan, Huiying Liang, and Ran Yang. 2018.
Classification of Bacterial and Viral Childhood Pneumonia Using Deep

1382

Authorized licensed use limited to: UNIVERSITY OF NEW MEXICO. Downloaded on May 16,2021 at 14:11:50 UTC from IEEE Xplore. Restrictions apply.