Deep learning based Diagnosis of diseases using Image

Classification
2021 11th International Conference on Cloud Computing, Data Science & Engineering (Confluence) | 978-1-6654-1451-7/20/$31.00 ©2021 IEEE | DOI: 10.1109/Confluence51648.2021.9377154

Shiva Singh Tomar

Ansh Saxena Gaurav Jain
Department of Computer Science and
Department of Computer Science and Department of Computer Science and
Engineering, Amity School of
Engineering, Amity School of Engineering, Amity School of
Engineering and Technology, Amity
Engineering and Technology, Amity Engineering and Technology, Amity
University Noida.
University Noida. University Noida.
New Delhi, India
New Delhi, India New Delhi, India
[email protected]
[email protected] [email protected]
Richa Gupta
Department of Computer Science and
Engineering, Amity School of
Engineering and Technology, Amity
University Noida.
New Delhi, India
[email protected]

Abstract— Any disease whether it is curable or not, it must are some symptoms like Cough, Fever as Most of the people
be diagnosed properly with some time in hand to take the with this condition suffer from sudden fever, Shortness of
appropriate actions in time. As it is popularly said that early Breath because the air sacs within the lungs can fill with fluid
detection of any disease is half cured. For detection of the or pus if you have got this condition which also causes a
diseases like Pneumonia, Covid-19, Brain Tumor Radiography,
Computed Tomography are a technique popularly used
cough but also difficulty in breathing and a pain because the
nowadays. The motivation towards the study of the topic was infection within the lungs can cause pain in elders while
that due to our country’s population density, the vulnerability coughing and breathing. The pain is chronic and may be
of getting infected and not being treated nicely was exposed extremely painful.
during the outbreak of Covid-19. The ratio of doctors in India is In today’s world, pneumonia is that the very leading
nearly 1:1456 which means 1 doctor has approximately 1456 explanation for death in children and one among the highest
patients to deal with. That also results in a lot of time being diseases with such mortality rates worldwide. It is a sort of a
wasted on diagnosis, scheduling appointments, collection of lung infection that is caused due to the viruses, fungi, or
reports, etc. which could prove to be critical for a patient. To
bacteria. The detection of pneumonia is usually administered
reduce all the time being wasted, with the help of machine
learning we intend to learn if we can predict if the patient is
by examining chest X-Ray radiography by highly trained
infected with a certain disease or not. How do we do that is by specialists. the method is extremely tedious, first getting the
using deep learning models and analyze on the basis of a few X- appointments from the highly trained doctors then their
Ray scan of the Chest to detect Pneumonia. availability causes delays in detection and their treatment.
Often, the method of detection also results in a disagreement
Keywords—Chest X-Ray Scans, Pneumonia, CNN- between radiologists. For detection of the diseases like
Convolutional Neural Networks, Deep Learning, Transfer Pneumonia, Computed Tomography are a technique
Learining popularly used nowadays. The CXRs which stand for chest
X-rays are very simple, economical because of which also the
I. INTRODUCTION most commonly used tool for diagnosing lung diseases.
When any experienced radiologist interprets an X-ray it is
Any disease whether it is curable or not, it must be diagnosed diagnosed as either normal or perceive to have a disease like
properly with some time in hand to take the appropriate carcinoma, tuberculosis, or pneumonia. one among the
actions in time. As it is popularly said that early detection of foremost common chest diseases as pneumonia, a lung
any disease is half cured. As we age our body’s natural infection caused by viruses, bacteria, or fungi. We have
defenses subsided reliable and, as a result, seniors are more observed that the computer-aided diagnosis systems have
vulnerable to infection – including pneumonia. this is often shown the potential for improving diagnostic accuracy. We
an illness that seems to affect the elderly during a radically also intend to perform experimental studies that further
different way than how it does to people in other age groups. demonstrate that this dataset is useful for Developing organ-
The age brings several health challenges but there are a wise models to detect various types of diseases based on
couple of health complications, which will be taken care of Image Classification. Image and the part of the body given as
by recognizing and identifying the first signs. one among input and given to the classifier which in-turn gives it to the
them is pneumonia. As we age our body’s natural defenses model and it classifies the image as the detection of disease,
subsided reliable and, as a result, seniors are more vulnerable or it is a normal condition. Therefore, we suggest the use of
to infection – including pneumonia. this is often an illness an automated analysis of chest CT images for the detection of
that seems to affect the elderly during a radically different Pneumonia. It helps in reducing the critical time before
way than how it does to people in other age groups. deciding the appropriate course of action. However, it does
Pneumonia within the elderly is often life-threatening. There not recommend the removal of confirmatory tests like Blood

399
© IEEE 2021. This article is free to access and download, along with rights for full text and data mining, re-use and analysis.
Authorized licensed use limited to: IEEE Xplore. Downloaded on February 17,2025 at 17:57:37 UTC from IEEE Xplore. Restrictions apply.
 tests which are utilized to confirm the infection in a person augmentations were used to reduce overfitting. Some of the
and to undertake to spot the sort of organism which is causing utilized augmentations were like scaling to a level of 512 x
the infection. Pulse oximetry is also used commonly to check 512 px resolution because they felt that the 256 resolution
the oxygen level in the patient’s blood. What Pneumonia could yield degradation of the results, mild rotations, shifting,
actually does which proves to be fatal is its ability to prevent scaling, shearing, and the horizontal flip. For a set of images,
the lungs from moving enough oxygen into the patient’s the research gave them even a random level of blur, noise, and
bloodstream. Nowadays, even the Sputum test are utilized brightness which completes the section of augmentations that
which takes a sample of fluid from the patient’s lungs were utilized.
(sputum) which is taken after a deep cough and then analyzed Then the dataset was split into a training dataset which
to assist explanation for the infection. included data for 25684 patients 27 and the testing set which
So, we wanted to figure towards it so on contribute our work had data for 1000 patients. During the process of model
towards society and help within the early diagnosis of the training, 28 they had utilized a range of base models which
diseases so on help people in getting the required treatment were pre-trained on the ImageNet dataset 24 incorporating the
as soon as possible which in-turn saves lives too. transfer learning.

II. RESEARCH FINDINGS They have tested a number of different encoder

architectures like NASNet-A-Mobile, Xception, PNASNet-5-
A. Datasets Used Large Inception-ResNet-v2, SE-ResNext-50, -101, ResNet-
34, - 50, -101 , and DualPathNet92. The results show a lesser
1. The ChestX-ray14 dataset which was publicly validation loss for SE_ResNext-101 after 18 epochs on the
released by Wang [1]. The dataset contains dataset. The only validation loss is mentioned, and nothing has
approximately 112,120 frontal-view X-ray images of been said about the training accuracy, testing accuracy,
as many as 30,805 unique patients. For the pneumonia precision, or recall of the models they created. They have also
detection task, they randomly split the dataset into not told how efficient the training process was. We opine that
training which has the data of 28744 patients having the training must have been a very hectic task as it involves
around 98637 images, validation set containing data over 25000 images in their dataset.
of 1672 patients and nearly 6351 images, and the test
data containing data of 389 patients, 420 images.
2. The dataset that has been utilized has 3 folders which
is segregated for the testing purposes, training
purposes along with validation purposes which are
further sub-divided into 2 folders one containing
pneumonia and the other with normal x-ray images,
respectively. The number of images used are 5,856 X-
beam pictures of front-back chests. This is a subset of
[1].
3. The dataset was publicly made available by releasing
it by the esteemed Radiological Society of the North
America on a very popular platform Kaggle.com. The
data containing x-ray images are intended to identify
if pneumonia is present in the x-ray data. The dataset
contains around 3.4 thousand images which is to be
utilized for the set of image training and
approximately the rest 1.2 thousand images for the set Figure 1 Sample Output of ChexNet Model
of image testing.
The authors Pranav Rajpurkarm, Andrew Y. Ng and
4. The Chest X-Ray image dataset hosted on many others [4], in their research work used the ChestX-
Mendeley.com [2]. The dataset had mainly two ray14 dataset which was publicly released by Wang [1]. They
subfolders which had the images belonging to downscaled the images to the resolution of 224×224 which
pneumonia (P) and normal (N) labelled images of the were then normalized based on the mean, standard deviation
chest X-ray, respectively. It had a total of five of the wholesome images in the training setoff the ImageNet.
thousand 8 hundred fifty-six X-ray images. CheXNet consisted of 121-layered network also called
convolutional neural network which was trained on the dataset
B. Literature ChestX-ray14, which was also the largest publicly available
According to the authors Gabruseva, Tatiana, Dmytro dataset of chest Xray that contains over a lakh frontal-view
Poplavskiy, and Alexandr Kalinin [3], they found it best to combined X-ray images with the frequently occurring 14
use the labeled dataset provided by the United States National diseases. They train their ChexNetmodel using the batch of
Institutes of Health Clinical Center which had the frontal view size 16.
Chest-X-Ray images from around 28000 unique patients. The
They have seen that the Dense-Nets improve the surge of
labels used were “Normal” signifying normal condition of the
patient’s lungs,” No Lung Opacity” signifying a situation data, gradients in the network, which ultimately makes the
which was not Normal, and” Lung Opacity” which showed tracing and optimization of very deep neural networks
there are some opaque particles present in the lungs is shown possible. They then supplant the ultimate fully connected layer
in Table 1. The tasks of images pre-processing, and with one that retains a single factor yield where a function
popularly known as sigmoid nonlinearity is applied.

400 2021 11th International Conference on Cloud Computing, Data Science & Engineering (Confluence 2021)

Authorized licensed use limited to: IEEE Xplore. Downloaded on February 17,2025 at 17:57:37 UTC from IEEE Xplore. Restrictions apply.
 They have used an initial learning rate of 0.001 throughout the execution of this technique, what it does to the image is
the model training process which is decayed by a factor of 0.1 that the average Red, Green and Blue values are found and 38
every time the validation loss has a flat maximum and plateaus Then, all the values of RGB held originally are factored by
after an epoch, which helps in the picking of the model with multiplying with the already calculated and the mentioned
the lowest validation loss. they had an accuracy of merely average value for expansion of the color scheme which
0.7680 which was obviously a much better attempt in the increases the overall values of RGB and yields us a colorized
classification of diseases that uses the ChestX-ray14 dataset version of the image. For the network that has been
containing 112,120 frontal-view X-ray images of 30,805 constructed, they have used a three layered convolutional 31
unique patients. The model ie. CheXNet, is a 121-layer neural network and consisting of a max pooling layer. Along
convolutional neural network which has an average of 85% With the hyperparameter fine-tuning, they have inserted a
accuracy as shown in Table-2. Although the accuracy is dropout layer following the Softmax function for the problem
significant, the training time is around 4-5 hours with heavy of over-fitting and to 40 prevent it.
GPU requirements as it is a 121-layered model that takes
significant time to get trained. They have executed a convolutional neural network which
is followed by a residual neural network to be used for the
According to the research work done by researchers - Mrs. classification process in the network.
Pravallika Rudraraju, K.Usha, Kodali Sri Harsha,
S.Rajeswari, B.V.S.S.Madhuri [5], they have developed a The loss function utilized in the model was a Binary cross
CNN model using the sequential model. The dataset that has entropy function as 41 it is binary classificati0n problem. The
been utilized has 5,856 X-beam pictures of front-back chests models were trained over the training data for 120 11 epochs
out of which a total of 3,722 pictures were allotted to the that had a batch amount of 40 and a realizing rate of 0.001
training set and a few 134 pictures were taken to the testing which decreased on the plateau by a contributory factor of 0.1.
set to improve the accuracy. The augmentation used have been
shown in Table-3. The implementation is done through the
sequential model to build a network. Various methods in
sequential model are used to add the network layers like
Conv2D layer, Maxpooling layer, flatten layer, Fully
Connected or Dense layer which yields an accuracy of nearly
90.87%. Here, in this research paper there is no information
that was mentioned about what is the structure of the layers,
the training time, hardware requirements for the model to be
trained and the rudimentary scores like recall, precision and
loss as shown in Table-4.
The research work done by researchers Deniz Yagmur
Urey, Can Doruk Taktakoglu and others [6] proposed to go
for a two-step processing of the image before the dataset goes
in the training procedure in their model built incorporating
deep learning. For creating the characteristic features of an x-
ray image sharper and explicitly recognizable and for easing
the process of classification. They have executed a
convolutional neural network followed by a residual neural Fig. 2. X-Ray Image Processing
network for the classification process. The data in this
publication used has been released at a public site on the The results were satisfactory as the model with increased
internet namely kaggle.com by the Radiological Society of contrast was vital for getting the edge-like characteristics that
North America. Their experimental methodology mainly are sharper in the picture and it therefore had an accuracy of
involves 3 different techniques of image processing: which are 78.73% during the usage in the fine-tuned edition exceeding
increasing the contrast as the first one, widening of the image the earlier highest notching value from CheXnet.
color space as the second one and artificially making changes
in the lighting of the image (increase/decrease in brightness) According to the research work done by researchers -
making it to the third image processing technique. Okeke Stephen and Do-Un Jeong and several others [7],
their trials were based upon the chest X-Ray image dataset
For data pre-processing, the Increment in brightness is hosted on Mendeley.com [2]. The dataset had mainly two a
executed in a way by parsing through each individual pixel of subfolder that contains pneumonia (P) and the normal (N)
the image-set, which increases the value of their respective chest X-ray images, independently. It had a total of five
values of Red, Green and Blue by a constant. The second thousand 8 hundred fifty-six X-ray images. They had adopted
technique that has been utilized here, was the significant the Keras open-source deep learning structure with the
increment in image 39 contrast. This is like the changing of TensorFlow backend carry on a basic PC together with a GPU
image brightness. By Increasing the image contrast, what it unit being Nvidia GeForce GTX TITAN Xp GPU card of 12
does to the image is that it creates the edges in images much GB, and CUDA Toolkit 9.0. Out of the total 5856 number of
more solid and some regions become noticeably much more images in the dataset, a total of three thousand seven hundred
visible as it was in its color scheme present originally. As it twenty two images were assigned to the training set and the
doesn’t reflect the features, henceforth, it becomes even more rest of the two thousand one hundred thirty four images were
essential for emphasizing certain parts of the image. The third allocated to the validation set to enhance the validation the
technique was all about the expansion of the color scheme by precision significantly.
applying certain procedures utilizing image processing. After

2021 11th International Conference on Cloud Computing, Data Science & Engineering (Confluence 2021) 401

Authorized licensed use limited to: IEEE Xplore. Downloaded on February 17,2025 at 17:57:37 UTC from IEEE Xplore. Restrictions apply.
 Here as shown in the fig.3, Their model had the feature folders one containing pneumonia and the other with normal
extractors which comprised of a conv3 × 3, 32 layer, conv3 × x-ray images, respectively. The number of images used are
3, 64 layer, conv3 × 3, 128 layers, another conv3 × 3, 128 5,856 X-beam pictures of front-back chests. This is a subset
layers, a max-pooling layer of size 2 × 2, and RELU activator of [1]. The Chest X-Ray image dataset hosted on
function between them. The results they achieved were like Mendeley.com [2]. The dataset had mainly two subfolders in
the training loss being 0.1288, training accuracy being 0.9531, the the dataset which had bifurcation in pneumonia (P) and
validation loss being 0.1835, and validation accuracy of normal (N) images of chest X-ray, respectively. It comprises
0.9373 after 100 epochs which are quite good in comparison of a total of five thousand eight hundred fifty-six number of
to the other research works previously done in this field. The X-ray images.
drawback they had was of the input array that was being used
as 200*200*3 which increases the time for each epoch during Table – 1 The Representation of the distribution of Classes in
the training process and as the model architecture being used the dataset (0 – normal, 1 – pneumonia) [3]
has several CNN layers it requires very high computation
power and proves to be a time-consuming process as training
the CNN architecture for an input size of 200*200*3 over a Class Target Patients
dataset of nearly 5000 images on these many layers in the Lung Opacity 1 9555
model requires much computational power and nearly 8 hours No Lung Opacity /
of time. Not Normal 0 11821
Normal 0 8851

Table – 2 CheXNet [4]: Radiologist-Level Pneumonia

Detection on Chest-X-Rays results.

Wang
Yao et
et al
Fig. 3. Model Architecture Pathology al CheXNet
(2017)
(2017)
[4]
In the research done by the authors Dimpy Varshni,
Kartik Thakral, Lucky Agarwal, Rahul Nijhawan, Atelectasis 0.716 0.772 0.8094
Ankush Mittal [8], the dataset which is used is the most Cardiomegaly 0.807 0.904 0.9248
popular dataset ChestX-ray14 released by Wang Effusion 0.784 0.859 0.8638
[1]additionally available to the public on the Kaggle a
platform that is comprised of 112,120 anterior chest X-ray Infiltration 0.609 0.695 0.7345
images from 30,085 patients. Here, the original images Mass 0.706 0.792 0.8676
contained 3-channel images which were to resize starting at
Nodule 0.671 0.717 0.7802
1024×1024 incorporated in 224×224 pixels which decreases
the substantial computation required and which also results in Pneumonia 0.633 0.713 0.768
the faster processing for the project and the model. The model Pneumothorax 0.806 0.841 0.8887
training has been implemented around the pre-attained
reduced images to reduce complexity of the model. Consolidation 0.708 0.788 0.7901
Edema 0.835 0.882 0.8878
For the process of extraction of features, the definitive
feature interpretation received in the process were inferred as Emphysema 0.815 0.829 0.9371
a 50176×1 component vector which subsequently provided as Fibrosis Pleural 0.769 0.767 0.8047
input to the various classifiers. Then, meant for pre-trained
Thickening 0.708 0.765 0.8062
CNN models like Xceptions, VGG16s, VGG-19NET,
ResNet-50, DenseNet-121 and DenseNet-169, they’ve Hernia 0.767 0.914 0.9164
assessed their effectiveness followed closely by various
classification systems involving Naive Bayes, Random Forest,
Support Vector Machine (SVM) and K-nearest neighbors.
The findings that were obtained, signified that the CNN
model of ResNet-50 which was of depth 168 had SVM
classifier performed significantly well than everything the
previous pre-trained CNN models achieving an AUC notch of
0.7749.
Here the downfall was that only area under the curve was
mentioned and nothing was mentioned about the primitive
scores like the training accuracy, testing accuracy, precision
or recall scores as shown in Table-5.

C. Results and Discussion

The dataset that has been utilized has 3 folders which is
segregated for the testing purposes, training purposes along
with validation purposes which are further sub-divided into 2

402 2021 11th International Conference on Cloud Computing, Data Science & Engineering (Confluence 2021)

Authorized licensed use limited to: IEEE Xplore. Downloaded on February 17,2025 at 17:57:37 UTC from IEEE Xplore. Restrictions apply.
Table – 3 The Representation of the augmentations utilized
VGG-16 Nalve Bayes 0.6193
during the training process to reduce the overfitting [5].
k-nearest
VGG-16 0.6847
Method Setting neighbors
Random
VGG-16 0.6563
Rescale 1/255 Forest
SVM(rbf
Beam shift 0.2 VGG-19 0.5
kemel)
loftiness shift 0.2
VGG-19 Nalve Bayes 0.5952
Rotation range 40
k-nearest
VGG-19 0.68502
Horizontal flip FALSE neighbors
Random
VGG-19 0.6481
Forest
Table – 4 The Representation of the results obtained via SVM(rbf
different techniques during the experimentation process ResNet-50 0.7749
kernel)
where the CNN with Lightened Image on Increased Contrast
with ResNet gives the best results in the classification of ResNet-50 Nalve Bayes 0.6891
pneumonia used in the research work [5].
k-nearest
ResNet-50 0.7298
neighbors
Network Accuracy
Random
CheXnet 0.768 ResNet-50 0.5793
Forest
(previously
DenseNet- SVM(rbf k
proposed model) 0.7577
121 kernel)
CNN with 0.6374
Unmodified Input DenseNet-12 Nalve Bayes 0.6691
CNN with 0.6542 DenseNet- k-nearest
Expanded Color 0.6981
121 neighbors
Scheme DenseNet- Random
CNN with 0.6992 0.6771
121 Forest
Increased DenseNet- SVM(rbf
Contrast 0.7476
169 kernel)
CNN with 0.7565 DenseNet-
Lightened Image Narve Bayes 0.6758
169
on Increased DenseNet- k-nearest
Contrast 0.6835
169 neighbors
CNN with 0.7873 DenseNet- Random
Lightened Image 0.6733
169 Forest
on Increased
Contrast with D. CONCLUSION
ResNet Any disease whether it is curable or not, it must be diagnosed
properly with some time in hand to take the appropriate
actions in time. As it is popularly said that early detection of
Table – 5 The Representation of the Evaluation results of any disease is half cured. As we age our body’s natural
different algorithms while the classification of Pneumonia. It defenses subsided reliable and, as a result, seniors are more
clearly shows that the Support Vector Machine shows pretty vulnerable to infection – including pneumonia. this is often
good results when it comes to the solution to the problem of an illness that seems to affect the elderly during a radically
classifying pneumonia [8]. different way than how it does to people in other age groups.
In today’s world, pneumonia is that the very leading
Feature explanation for death in children and one among the highest
Classifier AUC diseases with such mortality rates worldwide. it's a sort of
Extractor
SVM(rbr lung infection caused by viruses, bacteria, or fungi. The
XCeption 0.7034 detection of pneumonia is usually administered by examining
kernel)
chest X-Ray radiography by highly trained specialists. the
XCeption Nalve Bayes 0.6362 method is extremely tedious, first getting the appointments
k-nearest from the highly trained doctors then their availability causes
XCeption 0.6867 delays in detection and their treatment.
neighbors
Random The merits of the papers discussed in this literature review are
XCeption 0.6406 that the papers are a very good foundation and the
Forest
SVM(rbf steppingstone for the researcher who want to research in the
VGG-16 0.5 field of image classification. Also, it gives an idea to the
kernel)

2021 11th International Conference on Cloud Computing, Data Science & Engineering (Confluence 2021) 403

Authorized licensed use limited to: IEEE Xplore. Downloaded on February 17,2025 at 17:57:37 UTC from IEEE Xplore. Restrictions apply.
 person reading them about the various technologies and ray Database and Benchmarks on Weakly-
various attempts at the problem statement which were Supervised Classification and Localization of
successful and could be further researched to gain better Common Thorax Diseases,” May 2017, doi:
success rates in future. 10.1109/CVPR.2017.369.
The demerits include that it needs more research work to get [2] D. Kermany, K. Zhang, and M. Goldbaum, “Large
better success rates using the different quantitative measures Dataset of Labeled Optical Coherence Tomography
and more study in this field. (OCT) and Chest X-Ray Images,” 2018.
Hereby, this study aimed to provide the platform to researcher [3] T. Gabruseva, D. Poplavskiy, and A. Kalinin, “Deep
in the field of image based medical diagnosis. In the article Learning for Automatic Pneumonia Detection,”
we have study the literature of major respiratory disease. The 2020. [Online]. Available:
result and discussion present the analysis of classification https://github.com/tatigabru/kaggle-rsna.
method for the patient dataset of Pneumonia and covid-19. [4] P. Rajpurkar et al., “CheXNet: Radiologist-Level
The reference for dataset from various sources are cited. The Pneumonia Detection on Chest X-Rays with Deep
early detection based on MRI and CT scan study is shown is Learning,” Nov. 2017, [Online]. Available:
the result section. The larger dataset plays important role in http://arxiv.org/abs/1711.05225.
the outcome of study. The evaluation metrics the different [5] P. Rudraraju, K. Sri Harsha, K. Usha, and
models used in the research paper. It shows the Comparison B.V.S.S.Madhuri, “Pneumonia Detection by X-Ray
of the accuracy utilizing different methods, different features, Images Using Deep Learning through CNN,”
and different feature extraction process. It shows DWT of 10- Journal of Xi’an University of Architecture &
fold cross validation shows the best possible accuracy. Technology, 2020.
In the research work studied, for pre-trained CNN models [6] C. J. Saul, D. Y. Urey, and C. D. Taktakoglu, “Early
like Xception, VGG16, VGG-19, ResNet-50, DenseNet-121 Diagnosis of Pneumonia with Deep Learning,” Apr.
and DenseNet-169, they have evaluated their performance 2019, [Online]. Available:
followed by different classifiers including Naive Bayes, http://arxiv.org/abs/1904.00937.
Random Forest, Support Vector Machine (SVM) and K- [7] O. Stephen, M. Sain, U. J. Maduh, and D. U. Jeong,
nearest neighbors. “An Efficient Deep Learning Approach to
Pneumonia Classification in Healthcare,” Journal of
ACKNOWLEDGMENT Healthcare Engineering, vol. 2019, 2019, doi:
The work was supported by Amity University, Noida, UP 10.1155/2019/4180949.
Department of Computer . [8] D. Varshni, R. Nijhawan, K. Thakral, A. Mittal, and
L. Agarwal, “Pneumonia Detection Using CNN
REFERENCES based Feature Extraction,” 2019.
[1] X. Wang, Y. Peng, L. Lu, Z. Lu, M. Bagheri, and R.
M. Summers, “ChestX-ray8: Hospital-scale Chest X-

404 2021 11th International Conference on Cloud Computing, Data Science & Engineering (Confluence 2021)

Authorized licensed use limited to: IEEE Xplore. Downloaded on February 17,2025 at 17:57:37 UTC from IEEE Xplore. Restrictions apply.