Lung Disease Detection and Classification with Deep Learning Approach

2021 36th International Technical Conference on Circuits/Systems, Computers and Communications (ITC-CSCC) | 978-1-6654-3553-6/21/$31.00 ©2021 IEEE | DOI: 10.1109/ITC-CSCC52171.2021.9501445

Araya Chatchaiwatkul1, Pasuk Phonsuphee2, Yurananatul Mangalmurti3,

Naruemon Wattanapongsakorn*
King Mongkut’s University of Technology Thonburi, Bangkok, Thailand
[email protected], [email protected], [email protected],
[email protected]*

Abstract detection/classification system. One of the most

popular techniques in AI is deep learning which is a
Nowadays, COVID-19 outbreak and respiratory model that consists of deep (many) layers and
symptoms globally take a huge number of people’s lives requires large computation time. It performs well
away. Especially, COVID-19, which is a pandemic when the dataset is sufficiently large. The deep
initially spreading out in the first quarter of the year learning technique that is commonly used in pattern
2020, heavily affects many people to die. Most countries
recognition is convolutional neural network (CNN).
have tried to find ways to solve and mitigate this outbreak
including respiratory diseases due to the mentioned Recent study on classification of chest X-ray
reason. We also face with insufficient number of medical image on lung diseases including Covid-19 has been
personnel and equipment to treat the diseases. The need published throughout 2020. In early May 2020,
of technology to analyze the images for the disease Tuncer et al. [1] considered various classification
detection is quite a challenge. In this work, we consider techniques to classify Covid-19 patients and normal
detecting and classifying many lung diseases from chest patients using chest X-ray images. In late May 2020,
X-ray images using a deep learning (artificial Khan et al. [2] used a deep learning technique which
intelligence) approach with VGG16 models. The lung is CNN to classify diseases using chest X-ray images.
diseases are COVID-19, Pneumonia and Pneumothorax. Another study applied a deep learning model on the
We use quite large published disease datasets. Our
X-ray and CT scan images. Their work used a CNN
detection and classification models give impressive
results providing between 93% and 100% accuracy, and made balance of the output classes [3]. Later in
precision, recall and F1-measure. Sept 2020, a different study used a transfer learning
on two CNNs but applied on small dataset [4].
Keywords: COVID-19, Pneumonia, Pneumothorax, In this paper, we use a convolutional neural
Deep learning, Lung disease detection. network to detect various lung diseases using X-ray
images. Large datasets are used for the classification
1. Introduction model development. The next section presents our
methodology with deep learning approach which is
In 2019, the world discovered a new lung disease a CNN (VGG16) model. Then section 3 discusses
COVID-19 or coronavirus. The disease turns the the dataset and parameter setting, followed by
world into pandemic phase. The disease causes experimental results and conclusion in sections 4
serious damage to lung and respiratory system. The and 5, respectively.
disease usually takes around 14 days before showing
the symptom. The method of diagnose COVID-19 2. Methodology: Deep Learning with CNN
takes around 2-4 hours before the patients know if (VGG16)
they are infected or not. There are many lung We implement the model by using the
diseases that can be detected by using chest X-ray convolutional neural network (CNN) concept. Also,
images. Since, the X-ray images of the infected we apply one of state-of-the-art models called
persons need to be diagnosed by radiologist. VGG16 to recognize the disease with chest X- ray
However, detecting the disease requires expertise images. We use this model as transfer learning and
from medical doctors. The need of technology to adding more neural network layers to fulfill the
analyze the images for the disease detection is quite model qualification. CNN
a challenge. Fig. 1 shows another CNN overview starting with
Chest X-ray images can be analyzed with input image which is a black cat until the output
artificial intelligent (AI) methods to make a smart layer. It is different from the previous CNN

Authorized licensed use limited to: SARDAR VALLABHBHAI NATIONAL INSTITUTE OF TECH. Downloaded on November 17,2024 at 09:20:06 UTC from IEEE Xplore. Restrictions apply.
 architecture since this figure 1 above shows conv increasing along the way as double and remain
layer and pooling layer more than one of each. stable at 512 channels. Furthermore, in VGG16, it
Apparently, pooling layer is put after conv layer applies max pooling after implementing
alternately three times, end with fully connected convolutional layers [7]. VGG16 architecture is
layer and output layer. Actually, it can have softmax quite the best performance and ever won the
regression before going to output layer, but this ILSVRC challenge in 2014 [8], but it has some
diagram does not show. Another thing that can disadvantages which is slow to train and spends a lot
notice is the feature extractor arranging from conv of disk memory in the computer as it has a high
layer 1 to conv layer 3. It conveys that since conv number of parameters and deep layers [7].
layer 1 to pooling layer 3 aim to extract feature that
exist in the input image. After pooling layer 3, it is B. Performance evaluation functions
in the process of classifier that aims to classify the About model evaluation, we evaluate the model
image into classes. In addition, if using softmax prediction performance by considering confusion
regression, the output will have multiple classes. If matrix, accuracy, precision, recall, and f1-measure.
using logistic regression, the output will be binary
classes. 3. Dataset and Parameter Settings

As a chest X-ray image reference, we brought all

chest X- ray images from Kaggle website which
contains many datasets everyone can access. Before
training these chest X-ray images, we gradually
manually delete unusable chest X-ray images
containing noise, some disturbance, medical tools
attached to the chest. Since these uncleaned images
affect the model performance, we have to remove
them out in order to get accurate model prediction.
Fig. 1. CNN overview [5]
Table 1. Chest X-ray images for each lung disease
Disease Chest X-ray images
COVID-19 3,042
Pneumonia 12,000
Pneumothorax 4,668
Normal 14,114
Table 1 shows the remaining chest X-ray images
for each disease. We are going to use most of them
in our experiments.
Fig. 2. VGG16 architecture
A. VGG16
VGG16 is one of the best models that have a high
number of parameters which is about 138 million
parameters [6]. It was proposed by K. Simonyan and
A. Zisserman [7]. Fig. 2 shows the structure of
VGG16 and how it processes when an image is
ready to pass at the leftmost side. The input image
size in VGG16 has to be 224x224x3 pixels.
Fig. 3. Cleaned and uncleaned chest X-ray images.
Convolutional filter is set to 3 by 3 with the stride of
1, and implements the same padding. The pooling Fig. 3 shows the COVID-19, Pneumonia,
used in VGG16 is max pooling which is set to be Pneumothorax, and Normal images between cleaned
2x2 with the stride of 2. and uncleaned characteristics. Each pair of chest X-
VGG16 architecture can be noticed that the size ray images at the upper left, upper right, lower left,
is reduced a half along the way because of the max and lower right are COVID-19, Pneumonia,
pooling process. Also, the number of channels is Pneumothorax, and Normal, respectively. We can

Authorized licensed use limited to: SARDAR VALLABHBHAI NATIONAL INSTITUTE OF TECH. Downloaded on November 17,2024 at 09:20:06 UTC from IEEE Xplore. Restrictions apply.
 notice that the uncleaned chest X-ray images have correctly classified 2380 images out of 2400 images,
the medical strings attached over the lung areas. and only 1 Normal X-ray image is misclassified out
These uncleaned chest X-ray images are then of 2400 images.
removed out and remain only the cleaned chest X-
ray images as input dataset for the model process. Table 3. Confusion matrix of Pneumonia and Normal
For each experiment, hyper-parameters include Predicted
learning rate, batch size, optimizer, activation Pneumonia Normal
function, image augmentation, number of epochs for Pneumonia 2399 1
training. These parameters are set and tuned to get Actual Normal 20 2380
the fitting model. In addition, we ensure the number
of chest X-ray images training in each experiment is
Case 3: COVID-19, Pneumonia, and Normal –
balanced.
Binary Class
4. Experimental Results and Discussion For the third experiment, the training set has
There are similar hyper-parameters used in each 7,200 chest X-ray images containing equally for
experiment. We set Adam as an optimizer, learning normal and abnormal classes. The validation set has
rate is 10−5, epochs is set to 30, mini-batch size is 2,400 chest X-ray images for both classes. Likewise,
128, activation function used is ReLU. The training, the test set has 2,400 images for both classes.
validation, and testing ratio is set to 60:20:20. We do Table 4. Confusion matrix of Normal and Abnormal
not apply image augmentation. The number of chest (COVID-19, Pneumonia, and Normal)
X-ray images using on each experiment is different Predicted
depending on the remaining number of each lung Abnormal Normal
disease chest X-ray image and what lung diseases 1181 19
Abnormal
are involved. For binary classification, we use a Actual
sigmoid activation function in the output layer. For Normal 20 1180
multiclass classification, we use a softmax Table 4 shows the 2x2 confusion matrix. We
activation function. can see that the normal chest X-ray images are
Case 1: COVID-19 and Normal – Binary Class correctly classified 1180 images out of 1200
For the first experiment, the training set has 3,600 images, and only 19 Abnormal (with disease) X-
chest X- ray images containing equally for COVID- ray images out of 1200 images are misclassified
19 and normal. The validation set has 1,200 chest X- as Normal.
ray images for both classes. Likewise, the test set has Case 4: COVID-19, Pneumonia, Pneumothorax,
1,200 images for both classes. and Normal – Binary Class
Table 2. Confusion matrix of COVID-19 and Normal For the fourth experiment, the training set has
Predicted
10,800 chest X-ray images containing equally
COVID-19 Normal normal and abnormal classes. The validation set
COVID-19 599 1 has 3,600 chest X-ray images for both classes.
Actual
5 595
The test set has 3,600 images for both classes.
Normal
Table 5. Confusion matrix of Normal and
Table 2 shows the 2x2 confusion matrix, where Abnormal (COVID-19, Pneumonia,
only 1 COVID-19 and 5 Normal X-ray images are Pneumothorax)
misclassified out of 1,200 images. Predicted
Case 2: Pneumonia and Normal – Binary Class Abnormal Normal
For this experiment, the training set has 14,400 Abnormal 1668 132
chest X-ray images containing equally for Actual
Normal 122 1678
Pneumonia and normal. The validation set has 4,800
chest X-ray images for both classes. Also, the test Table 5 shows the matrix, where the normal
set has 4,800 images for both classes. images are correctly classified 1678 out of 1800
images, and 132 Abnormal (with disease) images
Table 3 shows the 2x2 confusion matrix of out of 1800 images are misclassified as Normal.
pneumonia and normal binary class prediction. We
can see that the Pneumonia X-ray images are

Authorized licensed use limited to: SARDAR VALLABHBHAI NATIONAL INSTITUTE OF TECH. Downloaded on November 17,2024 at 09:20:06 UTC from IEEE Xplore. Restrictions apply.
 Case 5: COVID-19, Pneumonia, and Normal - two diseases which are COVID-19 and
Multiclass Pneumothorax has the overall result 93%. Lastly,
In this experiment, the training set has 5,400 the multi-class experiment tested among COVID-
chest X-ray images containing COVID-19, 19, Pneumonia, and normal gives slightly different
Pneumonia, and normal classes. The validation set results ranging from 93% to 98%.
has 1,800 images for these classes, while the test
set has 1,800 images. 5. Conclusion
Table 6. Confusion matrix of COVID-19, In conclusion, the convolution neural networks
Pneumonia, and Normal (multiclass) prediction
with VGG16 models are effectively and accurately
Predicted detect and classify many lung diseases from lung X-
COVID- ray images. This approach can be used to help many
Normal Pneumonia
19 people diagnose the suspected lung diseases by
Normal 587 4 9 themselves easily after obtaining their chest X-ray
Actual COVID-19 16 577 7 images. These models are parameter-free, so that
user without knowledge with deep learning can also
Pneumonia 18 5 577 use. We are currently implementing a web
Table 6 shows the matrix representing multi-class application with these detection/classification
classification, where most of COVID-19 images models so that people can get access with no charge.
(577 out of 600 images), as well as Normal (587 out
of 600 images) and Pneumonia (577 out of 600) References
images are correctly classified. [1] T. Tuncer, S. Dogan, and F. Ozyurt, “An Automated
Table 7. Disease detection results Residual Exemplar Local Binary Pattern and Iterative
Accura Precisi Recall Relief based COVID-19 Detection Method Using Chest
Cases cy on F1-score X-ray Image”, Chemometrics and Intelligent Laboratory
Case 1: Binary Class Systems, Vol. 203, 2020.
COVID-19 0.99 1.00 1.00 [2] A.I., Khan, J.L. Shah, and M.M. Bhat, “CoroNet: A
0.99 Deep Neural Network for Detection and Diagnosis of
Normal 1.00 0.99 0.99
COVID-19 from Chest X-ray Images”, Computer
Case 2: Binary Class
Methods and Programs in Biomedicine, Vol. 196, 2020.
Pneumonia 0.99 1.00 1.00 [3] A. Sedik, et.al., “Deploying Machine and Deep
Normal 1.00 1.00 0.99 1.00 Learning Models for Efficient Data-Augmented
Case 3: Normal & Detection of COVID-19 Infections”, Viruses, Vol. 12, No.
Abnormal 7, pp. 769, 2020.
Normal 0.98 0.98 0.98 [4] A. Abbas et al., “Classification of COVID-19 in Chest
0.98
Abnormal 0.98 0.98 0.98 X-ray Images Using DeTraC Deep Convolutional Neural
Case 4: Normal & Network”, The Inter J of Research on Intel Systems for
Abnormal Real Life Complex Problems, 2020.
Normal 0.93 0.93 0.93 [5] dshahid380, Convolutional neural network, Toward
Abnormal 0.93 0.93 0.93 0.93
Data Science, Feb. 25, 2019. Accessed on: Nov. 20, 2020.
Case 5: Multi-class
(COVID-19, [Online]. Available: https://towardsdatascience.com/
Pneumonia, Normal)
covolutional-neural-network-cb0883dd6529
Normal 0.95 0.98 0.96
[6] R. Thakur, Step by step VGG16 implementation in
COVID-19 0.97 0.98 0.96 0.97 Keras for beginners, Toward Data Science, Aug. 6, 2019.
Pneumonia 0.97 0.96 0.97 Accessed on: Nov. 25, 2020. [Online]. Available:
https://towardsdatascience.com/step-by-step-vgg16-
Table 7 shows summary of the disease detection
implementation-in-keras-for-beginners-a833c686ae6c.
results obtained from all 5 experimental cases. The [7] M. Hassan, VGG16 - convolutional network for
performance evaluation functions are accuracy, classification and detection, Neurohive, Nov. 20, 2018.
precision, recall and F-measure. For binary class of Accessed on: Nov. 25, 2020. [Online]. Available:
COVID-19 and Pneumonia, the performance https://neurohive.io/en/popular-networks/vgg16/
evaluation results range from 99% to 100%. For [8] pawangfg, VGG-16 CNN model, GeeksforGeeks,
normal and abnormal covering COVID-19 and Feb. 27, 2020. Accessed on: Nov. 25, 2020. [Online].
Available: https://www.geeksforgeeks.org/vgg-16-cnn-
Pneumonia, the overall result is 98%. Meanwhile, model/
another normal and abnormal experiment covering

Authorized licensed use limited to: SARDAR VALLABHBHAI NATIONAL INSTITUTE OF TECH. Downloaded on November 17,2024 at 09:20:06 UTC from IEEE Xplore. Restrictions apply.