International Conference on Emerging Technologies (ICET) 2022

Lung Disease Detection using Deep Learning

1st Syed Krar Haider Bukhari 2nd Labiba Fahad
Department of Computer Science (MS CS) Department of Computer Science (Associate Professor)
FAST National University of Computer and Emerging FAST National University of Computer and Emerging
Sciences, Islamabad, Pakistan Sciences, Islamabad, Pakistan
i202077@nu.edu.pk labiba.fahad@nu.edu.pk

Abstract—Lung diseases refer to many disorders affecting the
lungs, such as pneumonia, tuberculosis, lung cancer, and many
other breathing problems. Due to COPD (chronic obstructive
pulmonary disease), 3 million people die from this disease each
year and it is the third main leading cause of human death
worldwide. Early disease detection is important to take timely
preventive measures. These diseases can be identified through
CT and chest X-ray images of the lungs. The accurate detection
2022 17th International Conference on Emerging Technologies (ICET) | 978-1-6654-5992-1/22/$31.00 ©2022 IEEE | DOI: 10.1109/ICET56601.2022.10004651
Deep learning is the major trend in medical science in terms
of disease detection. It helps doctors make more accurate di-
agnoses of their patient’s health and suggest better treatments.
Deep learning uses artificial neural networks (ANN), recurrent
neural networks (RNN), and reinforcement learning models.
But convolutional neural networks (CNN) have made a big
difference in computer vision and image analysis.

of lung diseases is challenging due to high similarities between Lung disease detection is quite a challenging and demanding
different lung diseases and variations in a single disease. The
proposed work is focused on learning the less inter-class and high problem in medical, computer vision, and artificial intelli-
intra-class variations in different lung diseases. In the proposed gence. The intra-class variation in diseases that are difficult
approach, pre-processing, data augmentation and deep learning to categorize. These variations range in severity from minor
is used for classification of four categories including normal, to major. When visual changes are slight, such as when recog-
lung cancer, pneumonia, and tuberculosis. The representation nizing different diseases, the classification problem becomes
of minority classes is improved by using data augmentation
technique DCGAN. A comprehensive evaluation of the proposed more complicated. Inter-class differences are extremely minor
approach using a publicly available dataset demonstrates a better among these classes. As a result, identifying generic features
performance of the proposed approach compared to the existing for classification becomes more challenging.
approaches. Fig. 1 shows the images of these three diseases, which are
Index Terms—Pneumonia, Lung Cancer, Tuberculosis, Chest quite similar to each other, thus compromising the accuracy.
X-ray, DCGAN, Deep learning
Some classes have fewer images compared to others. If we
see the images of these three lung diseases through the naked
I. I NTRODUCTION eye, it is visually difficult to detect the disease because they
have less inter-class variation within two different classes and
Lung diseases like pneumonia, tuberculosis, and lung cancer they have high intra-class variation within the class, which
severely damage the lungs and cause breathing problems. Lung makes the detection and classification purposes more difficult
cancer begins in the lungs, and it is an abnormal growth and challengeable.
of cells in the tissues of the lungs. Lung cancer mostly
occurs in heavy smokers. The main cause of lung cancer
is smoking tobacco [1]. Pneumonia is a chronic disease that
may be present in one or both lungs. The air sacs filled with
fluid in the lungs cause cough, fever, and breathing issues.
The main causes of pneumonia are viral, fungal, or bacterial
infections [2]. Tuberculosis is caused by bacteria which is
called mycobacterium tuberculosis. A person can be affected
by pulmonary TB by sneezing and coughing droplets in the
air of an infected person. It can affect different parts of the
human body, like the brain and kidneys [3].
Due to COPD (chronic obstructive pulmonary disease),
three million people die each year worldwide [1]. Lung disease
detection at an earlier stage is very important to save the
Fig. 1. CT and X-ray images of lung diseases
patient from death. It can be detected through the images of
CT scans, X-rays, and MRI. Many diseases are now diagnosed To overcome this problem, we proposed an approach in this
using artificial intelligence, which has demonstrated its effi- research. The proposed work improved lung disease recogni-
ciency and great performance in automatic image classification tion accuracy by addressing both less inter- and high intraclass
problems through different deep learning approaches. variation. Moreover, improving the representation of minority
classes in the data set by using DCGAN.

978-1-6654-5992-1/22/$31.00 ©2022 IEEE

Proc-154

Authorized licensed use limited to: Trial User - Superior University. Downloaded on March 20,2023 at 11:01:45 UTC from IEEE Xplore. Restrictions apply.
 II. R ELATED W ORK
Chest X-ray and CT (computerized tomography) images are
easy for patients to get and low-cost, so they are used in
most countries to recognize lung diseases. The authors of [4]
used four different convolutional neural network models for
lung detection: ResNet-50, DenseNet-201, Inception-V3, and
VGG19. The performance evaluation metrics consisted of the
f1 score, precision, recall, and confusion matrix. However, the
approach is evaluated on a limited dataset. In [5], authors
have used deep learning models for diagnosing COVID-19,
pneumonia, and lung cancer. They used CT and x-ray images.
The CNN models were VGG19-CNN, ResNet152V2, and
ResNet152V2 for the detection. The authors of [6] used
convolutional neural network models to classify lung diseases,
but they only had a small volume of the dataset. Resnet,
VGG16, ResNet-50, and Inception-V3, which are pre-trained
networks, performed classification tasks by using the transfer
learning approach. They compared the performance of the
Shenzhen and Montgomery lung datasets.
In [7], the authors used CNN and ANN networks for
the detection. Their performance evaluation was based on
the following parameters: precision, accuracy, recall, and f1
score. In [8], authors have used CT images and improved
deep learning methods for the detection of lung cancer. They
used image segmentation, and their evaluation parameters
were precision, accuracy, recall, and f1 score. In [9], authors
have used four different convolutional neural network models;
Vgg16, Alexnet, DenseNet-201, and Google net for pneumo-
nia detection which were present in both lungs these images
of the chest X-rays were used in content-based image retrieval
techniques. In [10], the author has used the CNN model
for pneumonia disease detection. The chest x-ray images
were used for detection. VGG19-CNN, ResNet152V2, and
ResNet152V2 for the classification. The automatic detection
of the disease of TB bacilli from the microscopic sputum and
smear images by using the methods of deep learning [11].
In [12], authors have used deep CNN for the detection of
pneumonia-infection disease. They compare the performance
and finally, they concluded that VGG-19 performed better than
Inception V3. In [13], authors experimented with pneumonia
diseases detection by using the CNN a deep learning model;
VGG19-CNN, ResNet152V2, and ResNet152V2.
In [14], authors have used CNN as a deep learning model.
They used some chest x-ray images for classification. VGG19-
CNN, ResNet152V2, and ResNet152V2 were used for the de-
tection. The authors of [15] demonstrated deep learning tech-
niques using x-viral data sets. They proposed a CAD model for
the classification of diseases. For the evaluation metrics, they
use the ROC curve (AUC), sensitivity, and specificity. In [16]
and in [17], the authors performed experiments on lung disease
detection by using a microscopy image database (ZNSM-iDB)
dataset. For the feature pre-processing, a CNN approach was
followed from scratch. In [18], the authors have used the
Montgomery County (MC) pixels 3, Shenzhen 4, and CXR8
data sets for TB detection. They used deep learning techniques
Proc-155
Authorized licensed use limited to: Trial User - Superior University. Downloaded on March 20,2023 at 11:01:45 UTC from IEEE Xplore. Restrictions apply.
to develop a CAD-based tuberculosis diagnostic system. They
performed image masking techniques and segmentation tech-
niques. Evaluation metrics were precision, accuracy, recall,
and f1 Score. In [19], the authors performed experiments on
lung disease detection by using a data set of osteosarcomas.
Techniques like CAD, CAD x, and CNN were used. They
have performed image pre-processing steps for the feature pre-
processing. The performance evaluation metrics consisted of
the f1 score, precision, and recall.
In [20], the authors worked on pneumonia detection
by using COVID-19, pneumonia, and severity classification
datasets. They use transfer learning techniques; they use
CNN and ANN networks for detection. They performed data
augmentation for pre-processing. In [21], authors have used
operable lung cancer with chest radiographer data set for
detection They use the deep learning techniques CAD, CXRs,
DCNN, DLM, and DLML. For the feature pre-processing, the
proposed image resize, crop, and enhancement techniques are
used. The performance evaluation metrics consisted of the F1
score, precision, recall, and confusion matrix.
III. M ETHODOLOGY
In Existing approaches, either focus on pneumonia, lung
cancer, and TB individually. The existing work either fo-
cuses on less inter-class or high intra-class variations [1]–
[3]. However, very few studies addressed them together, so
we addressed both of these and achieved a high f1 score.
We have used three distinct settings to find the best possible
solution for our problem. First, we used the image pre-
processing techniques on the dataset to improve the features
of images for efficient utilization. Second, we use the GAN
(generated adversarial network) to balance the dataset, and
finally, we passes images into (CNN) convolution neural
networks for feature extraction. These four CNNs (DenseNet-
201, GoogleNet, ResNet-50, and MobileNet-V2) are used
with hyper-parameters to progress the classification. After
classification, we have diagnosed lung disease inferentially in
which class this disease lies.
Fig. 2. Block diagram of proposed method
A. Dataset
The dataset contains a large collection of CT and X-ray
images of normal, lung cancer, pneumonia, and tuberculosis.
Moreover, it contains imbalanced data. The data was gathered
from different sources [1], [3], [5]. The Pneumonia contains
5,863 chest x-ray images [5]. Lung cancer contains 10,000
chest CT images [1]. TB contains over 1000 chest X-ray
images [3]. The normal images contained 7,533 and the total
images were 24,416. Pneumonia and TB were the minority
classes.The problem of biasness occurred in data due to
imbalance dataset, the model become more biased towards
majority classes which produces the poor results. To overcome
this problem data augmentation is used to balance the data to
achieve better accuracy.
B. Dataset pre-processing
To increase the detection accuracy the image pre-processing
steps were applied, like image enhancement, image transfor-
mation, random crops, resize, rescale, image flips, and rotation.
Data augmentation was performed on the dataset to balance
the data for minority classes. DCGAN (deep convolutional
generative adversarial network) was used for this purpose. The
1000 new instances were created for minority classes.
C. DCGAN
GANs are generated adversarial networks that are used to
teach a deep learning model to produce synthetic data using the
same training data distribution. DCGAN contains two different
models a discriminator and a generator. The generator generate
the synthetic images that are similar to training images which
are used to train the model. The discriminator discriminate
the images either fake or real images. In a DCGAN, both the
generator and the discriminator are layers that combine con-
volutional and convolutional-transpose operations. The result
will be an RGB image with a size of 3 x 64 x 64.
Fig. 3. DCGAN Algorithm [3]
D. CNN Architecture for Lung Disease Classification
We used four deep neural networks with distinct architec-
tures and kernel sizes.
ResNet-50 is a convolutional network with 50 layers. There
are five phases in all, each with a convolution and identity
block. We propose extracting features from the output of
ResNet-50’s last convolutional block. The Conv5 block pro-
duces a 7* 7 *2048 dimensional array, which is used as the
Proc-156
Authorized licensed use limited to: Trial User - Superior University. Downloaded on March 20,2023 at 11:01:45 UTC from IEEE Xplore. Restrictions apply.
FC-1000 layer’s input.
DenseNet-201 DenseNet-201 is a convolutional neural net-
work which contains 201 deep layers with dense block. It
utilizes dense connections between layers which are connected
with each other. Each layer receive information from all
preceding layers and move forward to next layer.
MobileNet-V2 MobileNet-V2 is an image classification and
mobile vision CNN architecture model. There are various
models, but what distinguishes it’s a 53 layers convolutional
neural network.
GoogleNet GoogLeNet is a deep CNN that has 22 lay-
ers. Most people refer to this as ”GoogLeNet” (Inception-
V1).GoogleNet has nine of these inception modules that fit
together in a straight line. It consists of 22 layers (27, including
the pooling layers). The FC layers were switched out at the end
of the architecture for a globally averaged pooling approach.
This method finds the average of each of the system’s feature
maps.
IV. I MPLEMENTATION
In the implementation, we used Python with libraries like
Pandas, NumPy, TensorFlow, Keras, PyTorch, SKLearn, etc.
These are all standard libraries used in deep learning to
perform a wide range of classifications for images.
A. Evaluation Methodology
For evaluation purposes, we split the dataset into three
parts: training, testing, and validation as shown in fig 3. For
making fair comparisons with existing approaches, we used
the same split ratio of 70:30 in the dataset. We have used 70%
for training, 20% for testing, and 10% for the validation set,
as shown in a pie chart. We have used the 5 cross-validation
for this purpose. This evaluation methodology is used to
find differences learned by using our approach and deep
learning model to compute good results. We also validate
our approaches by comparing them with existing approaches
using the same datasets.
B. Evaluation Parameters
The performance of the model was evaluated with the use
of the following metrics: accuracy, precision, recall, and f1
score.The formulas are given below
• Accuracy: It shows the overall performance of the
model. It works well when all classes have the same
weight.
Accuracy = TP + TN / TP +TN + FP + FN
• Precision: The ratio of the number of samples accurately
recognized as positive to the total number of samples
identified as positive (either correctly or incorrectly). It
measures whether the predictions are accurate.
Precision = TP / TP + FP
• Recall: The number of positive samples that were cor-
rectly labelled as positive to the total number of positive
 samples.The more positive samples that are found, the C. Implementation of DCGAN
higher the recall.
Recall = TP / TP + FP In the implementation of DCGAN, each learning rate for
• F1-Score: The mean of precision and recall is the f1 optimizer was 0.005, the number of workers for data loader
score. To put it another way, the F1 score shows how was 2, the batch size of 64 training images was 300, and the
well the accuracy and recall work together. number of epchos was 300. The synthetic images produced
F1 Score = 2 * (Precision * Recall) / (Precision + Recall) using dcgan are much more similar to the real images, as
shown in Fig. 4.
The 1000 new instances were created to balance the
V. E XPERIMENTS AND R ESULTS dataset. After balancing the dataset, we implemented the
pre-processing steps like image enhancement and transforms.
In this section, we have discussed the details of the These processed images are passed to the CNN models for
experimental setup that was used in baseline models. We feature extraction and classification.
have conducted the initial experiment and implemented the
four CNN models: ResNet-50, DenseNet-201, GoogleNet, and
MobileNet-V2. First, we retrieved results from the dataset
without image pre-processing, then we applied image pre-
processing steps to the dataset and extracted the following
findings. By using the loss function sparse categorical cross-
entropy with SGD optimizer. We trained the dataset on 50
epochs and adjusted the parameters.

A. Results with Image Pre-Processing and without Image Pre-

Processing Technique
Table 1 shows the results for both. The results are obtained
without image pre-processing and are the baseline results.
These results are obtained by passing the dataset to proposed Fig. 4. Results of DCGAN
deep learning models. For image pre-processing techniques,
the first method consisted of randomly cropping a 224*224
pixel image. Second, flipped the image horizontally, which
allowed information regarding reflection invariance to be D. Generator and Discriminator loss
captured. Then finally, we used image enhancement. The
The loss of generator and discriminator is reduced at every
CNN models were Resnet-50, DenseNet-201, GoogleNet, and
iteration. It becomes stable after 4,000 iterations. No more
MobileNet-V2.
improvement in the loss is observed as shown in Fig. 5.

TABLE I
R ESULTS USING IMAGE PRE - PROCESSING AND WITHOUT USING IMAGE
PRE - PROCESSING

CNN Models Without Image With Image

Pre-Processing Pre-Processing
(Accuracy) (Accuracy)
GoogleNet 84% 88%
ResNet-50 87% 91%
DenseNet-201 89% 93%
MobileNet-V2 83% 87%

B. Proposed Approach
In our proposed approach, we have used image pre-
processing and DCGAN (deep convolutional generated ad-
versarial networks) for improvement in results to get higher
accuracy and make fair comparisons with existing approaches.
After fine-tuning hyper-parameters to the ideal level, our
proposed approach, DenseNet-201, demonstrates a significant
improvement in outcomes.
Fig. 5. Generator and discriminator loss
The number of workers for the data loader was 2. The batch
size was 64. All images will be resized to this 64. The number
of channels for colour images is 3 during the training phase.
The size of the z latent vector was 100. The size of the feature
maps was 64 in the generator and discriminator.

Proc-157

Authorized licensed use limited to: Trial User - Superior University. Downloaded on March 20,2023 at 11:01:45 UTC from IEEE Xplore. Restrictions apply.
E. DenseNet-201 TABLE II
R ESULTS USING DCGAN AND I MAGE P RE -P ROCESSING
In this approach, we have implemented the CNN architec-
CNN Models Accuracy Precision Sensitivity Recall F1
ture of denesenet-201, which contains 201 layers of dense Score
blocks. For training the images, the Adam optimizer was used GoogleNet 91% 0.93 0.89 0.84 0.91
with a batch size of 64, a learning rate of 0.002, and for the ResNet-50 94% 0.95 0.96 0.94 0.94
loss function, cross-entropy loss was used. The image size was DenseNet-201 96% 0.98 0.95 0.94 0.96
MobileNet-V2 93% 0.94 0.90 0.92 0.93
64*64. With 0.98 precision, 0.95 sensitivity, 0.94 recall, and
0.96 F1-score, we achieved 96% accuracy.
In Figure 6, we can see the training and testing loss are
decreasing in each epoch. VI. C OMPARISON WITH D IFFERENT M ODELS
In this section, comparisons were presented. Fig. 7, shows
the comparisons of four different deep learning models like
ResNet-50,DenseNet-201,MobileNet-V2 and GoogleNet. The
blue colour shows the f1 score of pneumonia, the orange
colour shows the f1 score of lung cancer, and the grey colour
shows the F1 score of tuberculosis. We compared the f1
score of each disease individually. Pneumonia gives 96% and
95% f1-score using DenseneNet-201 and ResNet-50. The
overall DenseNet-201 performs well by achieving 96.6% at
pneumonia. Lung cancer’s highest f1 score is 90%, and TB’s
highest f1 score is 91%. These results were obtained after
data augmentation, image pre-processing, and using DCGAN.

Fig. 6. DenseNet-201 train and test loss

In Figure 7, we can see the training and testing accuracy is

improving in each epoch.

Fig. 8. Individual disease comparison on the basis of F1 score

VII. C OMPARISON OF P ROPOSED A PPROACH WITH

Fig. 7. DenseNet-201 train and test accuracy
F. Result Using DCGAN
In our final approach DenseNet-201 demonstrate a signif-
icant improvement in outcomes. We achieved 96% percent
accuracy.
E XISTING A PPROACH
Table 3 shows the comparisons of the existing approach
with our approach. Existing approaches, either focus on
pneumonia, lung cancer, and TB individually. The existing
work focuses on fewer inter-class or high intra-class
variations [1]–[3]. However, very few studies addressed them
together, so we addressed both of these and achieved a high f1
score. It shows the comparison based on the f1 score of each
disease like pneumonia, tuberculosis, and lung cancer using
DenseNet-201. The highest f1-score achieved by pneumonia
diseases is 0.96.

Proc-158

Authorized licensed use limited to: Trial User - Superior University. Downloaded on March 20,2023 at 11:01:45 UTC from IEEE Xplore. Restrictions apply.
 TABLE III
R ESULTS C OMPARISON WITH P ROPOSED A PPROACH
Classes Existing Approach Proposed Approach
(F1-Score) (F1-Score)
Pneumonia [5] 0.90 0.96
Lung Cancer [1] 0.85 0.92
Tuberculosis [3] 0.87 0.90
VIII. C ONCLUSION AND F UTURE W ORK
In this study, we have successfully improved the clas-
sification accuracy of lung diseases by addressing the less
inter and high intra-class variation. To increase the detection
accuracy, the image pre-processing steps were applied, like
image enhancement, image transformation, random crops, re-
size, rescale, image flips, and rotation. Moreover, we improved
the representation of minority classes in the data set by using
a data augmentation technique known as DCGAN. In data
augmentation, the new instances were created to balance the
minority classes in the dataset. Our proposed approach to
DenseNet-201 significantly improved outcomes by achieving
96 % accuracy and an F1 score of 0.96. Furthermore, we com-
pared the performance of four CNN networks on the diagnosis
of pneumonia, lung cancer, and tuberculosis diseases. The
test results showed that DenseNet-201 outperforms on medical
images and classifies diseases very well. We performed three
experiments. In the first, we improved the recognition accuracy
of three lung diseases (pneumonia, lung cancer, and TB) by
addressing less inter and intra class variation. Secondly, we
have improved the representation of minority classes in the
data set by using DCGAN. Finally, we have compared the
proposed approach with state-of-the-art existing techniques to
show the effectiveness of our approach.
This work will be extended in the future to include more
specific methods for classifying lung diseases by increasing
the number of classes in the dataset. Also, we might improve
the results with the help of a pre-trained network trained on a
real-time dataset taken from hospitals’ chest x-ray images for
better classification and more accurate disease predictions.
R EFERENCES
[1] X. Wang, Z. Yu, L. Wang, and P. Zheng, “An enhanced priori knowledge
gan for ct images generation of early lung nodules with small-size
labelled samples,” Oxidative Medicine and Cellular Longevity, vol.
2022, 2022.
[2] L. Račić, T. Popović, S. Šandi et al., “Pneumonia detection using
deep learning based on convolutional neural network,” in 2021 25th
International Conference on Information Technology (IT). IEEE, 2021,
pp. 1–4.
[3] Z. I. Farooq and S. Srivastava, “Detection of tuberculosis by image
processing using gan’s.”
[4] Y. Gu, X. Lu, L. Yang, B. Zhang, D. Yu, Y. Zhao, L. Gao, L. Wu, and
T. Zhou, “Automatic lung nodule detection using a 3d deep convolutional
neural network combined with a multi-scale prediction strategy in chest
cts,” Computers in biology and medicine, vol. 103, pp. 220–231, 2018.
[5] E. Ayan and H. M. Ünver, “Diagnosis of pneumonia from chest x-ray
images using deep learning,” in 2019 Scientific Meeting on Electrical-
Electronics & Biomedical Engineering and Computer Science (EBBT).
Ieee, 2019, pp. 1–5.
Proc-159
Authorized licensed use limited to: Trial User - Superior University. Downloaded on March 20,2023 at 11:01:45 UTC from IEEE Xplore. Restrictions apply.
[6] J. X. Qiu, H.-J. Yoon, P. A. Fearn, and G. D. Tourassi, “Deep learning
for automated extraction of primary sites from cancer pathology reports,”
IEEE journal of biomedical and health informatics, vol. 22, no. 1, pp.
244–251, 2017.
[7] W. Khan, N. Zaki, and L. Ali, “Intelligent pneumonia identification from
chest x-rays: A systematic literature review,” IEEE Access, vol. 9, pp.
51 747–51 771, 2021.
[8] S. Tammina, “Covidsort: Detection of novel covid-19 in chest x-ray
images by leveraging deep transfer learning models,” in ICDSMLA 2020.
Springer, 2022, pp. 431–447.
[9] M. Zak and A. Krzyżak, “Classification of lung diseases using deep
learning models,” in International Conference on Computational Sci-
ence. Springer, 2020, pp. 621–634.
[10] G. A. P. Singh and P. Gupta, “Performance analysis of various machine
learning-based approaches for detection and classification of lung cancer
in humans,” Neural Computing and Applications, vol. 31, no. 10, pp.
6863–6877, 2019.
[11] T. Rajasenbagam, S. Jeyanthi, and J. A. Pandian, “Detection of pneumo-
nia infection in lungs from chest x-ray images using deep convolutional
neural network and content-based image retrieval techniques,” Journal
of Ambient Intelligence and Humanized Computing, pp. 1–8, 2021.
[12] M. El-Melegy, D. Mohamed, and T. ElMelegy, “Automatic detection of
tuberculosis bacilli from microscopic sputum smear images using faster
r-cnn, transfer learning and augmentation,” in Iberian Conference on
Pattern Recognition and Image Analysis. Springer, 2019, pp. 270–278.
[13] A. U. Ibrahim, M. Ozsoz, S. Serte, F. Al-Turjman, and P. S. Yakoi,
“Pneumonia classification using deep learning from chest x-ray images
during covid-19,” Cognitive Computation, pp. 1–13, 2021.
[14] H. Farhat, G. E. Sakr, and R. Kilany, “Deep learning applications in
pulmonary medical imaging: recent updates and insights on covid-19,”
Machine vision and applications, vol. 31, no. 6, pp. 1–42, 2020.
[15] S. S. Meraj, R. Yaakob, A. Azman, S. Rum, A. Shahrel, A. Nazri, and
N. F. Zakaria, “Detection of pulmonary tuberculosis manifestation in
chest x-rays using different convolutional neural network (cnn) models,”
Int. J. Eng. Adv. Technol.(IJEAT), vol. 9, no. 1, pp. 2270–2275, 2019.
[16] J. Zhang, Y. Xie, G. Pang, Z. Liao, J. Verjans, W. Li, Z. Sun, J. He,
Y. Li, C. Shen et al., “Viral pneumonia screening on chest x-rays using
confidence-aware anomaly detection,” IEEE transactions on medical
imaging, vol. 40, no. 3, pp. 879–890, 2020.
[17] N. Ali, S. Ansari, Z. Halim, R. H. Ali, M. F. Khan, and M. Khan,
“Breast cancer classification and proof of key artificial neural network
terminologies,” in 2019 13th International Conference on Mathematics,
Actuarial Science, Computer Science and Statistics (MACS). IEEE,
2019, pp. 1–6.
[18] S. Kant and M. M. Srivastava, “Towards automated tuberculosis de-
tection using deep learning,” in 2018 IEEE Symposium Series on
Computational Intelligence (SSCI). IEEE, 2018, pp. 1250–1253.
[19] Z. Yousuf, U. Akram, A. Jameel, and W. Raza, “Analysis of tuberculosis
detection techniques using chest x-rays: A review,” in 2018 IEEE
Conference on Systems, Process and Control (ICSPC). IEEE, 2018,
pp. 23–28.
[20] D. Anisuzzaman, H. Barzekar, L. Tong, J. Luo, and Z. Yu, “A deep
learning study on osteosarcoma detection from histological images,”
Biomedical Signal Processing and Control, vol. 69, p. 102931, 2021.
[21] M. La Salvia, G. Secco, E. Torti, G. Florimbi, L. Guido, P. Lago,
F. Salinaro, S. Perlini, and F. Leporati, “Deep learning and lung
ultrasound for covid-19 pneumonia detection and severity classification,”
Computers in biology and medicine, vol. 136, p. 104742, 2021.