EFFECTIVE MODEL FOR PNEUMONIA DETECTION FROM CHEST X-RAYS USING DEEP
CONVOLUTIONAL NEURAL NETWORKS
BY
NASEEM ANSARI
A dissertation submitted in fulfilment of the requirement for the degree of Master of Science (Electronics Engineering)
Kulliyyah of Engineering
International Islamic University Malaysia
NOVEMBER 2020
i
ABSTRACT
Pneumonia is a disease which occurs in the lungs due to bacterial, viral or fungal infection and causes lung alveoli to fill with pus or fluid. Chest x-ray is the most common diagnostic tool for pneumonia. However, because of several other conditions in the lungs such as volume loss, bleeding, fluid overload, lung cancer or post-radiation or surgical changes, the diagnosis of pneumonia in chest x-rays becomes very complicated. Therefore, there is urgent need for computer aided diagnosis systems to assist clinicians in making better decisions. In this work, a deep convolutional neural network, ResNet-50 architecture, is proposed and is trained using transfer learning technique. A pre-trained model on ImageNet dataset is used and with the use of transfer learning, ResNet-50 model is trained for binary classification of chest x-ray images into pneumonia and non-pneumonia. Two datasets have been used and the ResNet-50 model was implemented on both the datasets. The model achieved an accuracy of 96.76% with RSNA dataset and 94.06% with Chest X-ray Image (CXI) dataset. RSNA dataset despite having almost five times more images than CXI dataset took very less time for training. Also because of the use of transfer learning technique both the ResNet-50 models were able to learn the significant features of pneumonia with only 50 % training.
The proposed ResNet-50 model gave an accuracy of 96.76 %, however, the model can be improved by using more deeper networks. Furthermore, this work could be extended to detect and classify x-ray images consisting of both lung cancer and pneumonia.
It is verified.
(Muhammad Ibn Ibrahimy)
ii
ثحبلا ةصلاخ
ءلاتما ببسيو ةيرطف وأ ةيسويرف وأ ةييرتكب ىودع ببسب ينتئرلا في ثديح ضرم يوئرلا باهتللاا وأ ديدصلاب ةيوئرلا تلاصيولحا لئاوسلا
اًعويش رثكلأا ةيصيخشتلا ةادلأا يه ردصلل ةينيسلا ةعشلأا .
.يوئرلا باهتللال ببسب ،كلذ عمو
دوجو نم ديدعلا ىرخلأا ةئرلا ضارمأ
ضارعلأا تاذ ،ةبهاشلما
لثم
،فيزنلاو ،مجلحا نادقف ةدئازلا لئاوسلاو
وأ ةئرلا ناطرسو ، ةئرلا في تايرغتلا
لا وأ عاعشلإا دعب ام تايلمع
.ةياغلل اًدقعم ردصلل ةينيسلا ةعشلأا في يوئرلا باهتللاا صيخشت حبصي ،ةيحارلجا ةجاح كانه ،كلذلف
لأا ةدعاسلم رتويبمكلا ةدعاسبم صيخشتلا ةمظنلأ ةحلم فىو.لضفأ تارارق ذاتخا في ءابط
اذه ثحبلا دق ،
ةكبش حاترقا تم
فافتلإ ةقيمع ةينوبصع
Deep Convolutional Neural Network ،
ةينب
ResNet-50 ةينقت مادختساب اهبيردت تمو
ملعتلا لقن Transfer Learning
. جذونم لمعتساو
تانايب ةعوممج ىلع اًقبسم بَّردُم ImageNet
ت مادختسابو ق
ملعتلا لقن ةين جذونم بيردت تم ،
ResNet-50 لىإ ةينيسلا ةعشلأاب ردصلا روصل يئانثلا فينصتلا ىلع
باهتلإ باهتللاا يرغو ةئرلا
.يوئرلا تانايب تيعوممج مادختسا ّتمو ،
و جذونم ذّفن دق ResNet-50
.تانايبلا تيعوممج ىلع اذه
دقو
تغلب ةقد جذومنلا قّقح 96.76
تانايب ةعوممج عم ٪ RSNA
و 94.06 تانايب ةعوممج عم ٪
( ةينيسلا ةعشلأاب ردصلا ةروص CXI
تانايب ةعوممج تقرغتساو .) RSNA
اتقو ىلع بيردتلل لقأ
تانايب ةعوممج نم اًبيرقت تارم سمبخ رثكأ روص دوجو نم مغرلا ةينقت مادختسا ببسب اًضيأو . CXI
قن ملعتلا ل يزارط نم لك نّكتم ،
ResNet-50 باهتللال ةمهلما تامسلا ملعت نم
هبيردتب يوئرلا
ىلع 50 تانايبلا نم طقف%
. جذونم ىطعأف Resnet-50
غلبت ةقد حترقلما 96.76
،كلذ عمو ،٪
.اًقمع رثكأ تاكبش مادختساب جذومنلا ينستح نكيم لمعلا اذه عيسوت نكيم ،كلذ لىإ ةفاضلاابو
اشتكلا تيلا ةينيسلا ةعشلأا روص فينصتو ف لك ىلع يوتتح
.يوئرلا باهتللااو ةئرلا ناطرس نم
iii
APPROVAL PAGE
I certify that I have supervised and read this study and that in my opinion, it conforms to acceptable standards of scholarly presentation and is fully adequate, in scope and quality, as a dissertation for the degree of Master of Science (Electronics Engineering)
………..
Muhammad Ibn Ibrahimy Supervisor
………..
S.M.A Motakabber Co-Supervisor
I certify that I have read this study and that in my opinion it conforms to acceptable standards of scholarly presentation and is fully adequate, in scope and quality, as a dissertation for the degree of Master of Science (Electronics Engineering)
………..
Noreha bt Abd. Malik
………..
Suriza bt Ahmad Zabidi
This dissertation was submitted to the Department of Electrical and Computer Engineering and is accepted as a fulfilment of the requirement for the degree of Master of Science (Electronics Engineering).
………..
Mohamed Hadi Habaebi Head, Department of Electrical and Computer Engineering
This dissertation was submitted to the Kulliyyah of Engineering and is accepted as a fulfilment of the requirement for the degree of Master of Science (Electronics Engineering).
………..
Sani Izan Ihsan
Dean, Kulliyyah of Engineering
iv
DECLARATION
I hereby declare that this dissertation is the result of my own investigations, except where otherwise stated. I also declare that it has not been previously or concurrently submitted as a whole for any other degrees at IIUM or other institutions.
Naseem Ansari
Signature ... Date ...
v
INTERNATIONAL ISLAMIC UNIVERSITY MALAYSIA
DECLARATION OF COPYRIGHT AND AFFIRMATION OF FAIR USE OF UNPUBLISHED RESEARCH
EFFECTIVE MODEL FOR PNEUMONIA DETECTION FROM CHEST X-RAYS USING DEEP CONVOLUTIONAL NEURAL
NETWORKS
I declare that the copyright holders of this dissertation are jointly owned by the student and IIUM.
Copyright © 2020 Naseem Ansari and International Islamic University Malaysia. All rights reserved.
No part of this unpublished research may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means, electronic, mechanical, photocopying, recording or otherwise without prior written permission of the copyright holder except as provided below
1. Any material contained in or derived from this unpublished research may be used by others in their writing with due acknowledgement.
2. IIUM or its library will have the right to make and transmit copies (print or electronic) for institutional and academic purposes.
3. The IIUM library will have the right to make, store in a retrieved system and supply copies of this unpublished research if requested by other universities and research libraries.
By signing this form, I acknowledged that I have read and understand the IIUM Intellectual Property Right and Commercialization policy.
Affirmed by Naseem Ansari
……..……….. ………..
Signature Date
vi
ACKNOWLEDGEMENTS
Alhamdulillahi Rabbil Aalmeen.
All praise to Almighty Allah who has given me good health, patience and support each and every day during my study and finally helped me complete my thesis.
Firstly, it is my utmost pleasure to dedicate this work to my dear parents and my family, who granted me the gift of their unwavering belief in my ability to accomplish this goal: thank you for your support and patience.
A special thanks to Professor Dr. Muhammad Ibrahimy and Associate Professor Dr. S.M.A Motakabber for their continuous support, encouragement and leadership, and for that, I will be forever grateful.
I would like to express my sincere gratitude to my fellow labmates especially Br. Rimaz who helped me a lot whenever I got stuck and helped me in the project completion.
Finally, I wish to express my appreciation and thanks to those who provided their time, effort and support for this project. To the members of my dissertation committee, thank you for sticking with me.
vii
TABLE OF CONTENTS
Abstract ... i
Abstract in Arabic ... ii
Approval Page ... iii
Declaration ... iv
Copyright Page ... v
Acknowledgements ... vi
Table of Contents ... vii
List of Tables ... x
List of Figures ... xi
List of Symbols ... xiv
List of Abbreviations ... xv
CHAPTER ONE: INTRODUCTION ... 1
1.1 Introduction... 1
1.2 Problem Statement ... 5
1.3 Research Objectives... 6
1.4 Research Methodology ... 7
1.5 Research Scope ... 8
1.6 Thesis Organization ... 8
CHAPTER TWO: LITERATURE REVIEW ... 10
2.1 Introduction to Deep Learning... 10
2.1.1 Deep Learning: Machine Learning inspired by Human Brain ... 10
2.1.2 Types of Deep Learning Algorithms ... 12
2.1.3 Convolutional Neural Networks ... 16
2.1.3.1 Convolution Layers ... 17
2.1.3.2 Pooling Layers ... 21
2.1.3.3 Fully Connected Layers ... 22
2.1.4 Training A Convolutional Neural Network ... 23
2.1.4.1 Different Kinds of Classification Problems ... 23
2.1.4.2 Training Data Sets ... 25
2.1.4.3 Activation Function for the Output Layer ... 26
2.1.4.4 Choosing the Cost Function ... 27
2.1.4.5 Training A CNN to Convergence ... 29
2.1.5 Evaluation Metrics ... 30
2.2 Deep Learning In Medical Imaging ... 36
2.2.1 From Traditional Machine Learning to Deep Learning ... 36
2.2.2 Related Work ... 39
2.2.2.1 Challenges ... 47
2.3 Basic Interpretation Case of Pneumonia in Chest X-Ray ... 49
2.3.1 Consolidation ... 49
2.3.2 Lung Abscess ... 53
2.3.3 Secondary Complications ... 55
2.4 Summary ... 56
viii
CHAPTER THREE: RESEARCH METHODOLOGY ... 57
3.1 Introduction... 57
3.2 General Structure of the Implementation ... 57
3.3 Framework ... 58
3.4 Dataset ... 59
3.4.1 RSNA Pneumonia Detection Challenge Dataset ... 59
3.4.2 Chest X-Ray Images Dataset ... 61
3.5 Pre-Processing ... 61
3.5.1 DICOM Data Extractor ... 62
3.5.2 Image Enhancement ... 63
3.5.3 Splitting the Dataset into Training, Validation and Test Set ... 66
3.6 Training the CNN Classifier ... 67
3.6.1 Transfer Learning ... 68
3.7 Classification Using Resnet-50 ... 70
3.8 Evaluation of the Model ... 74
3.9 Prediction ... 75
3.10 Summary ... 75
CHAPTER FOUR: RESULTS AND DISCUSSIONS ... 76
4.1 Introduction... 76
4.2 Simulation on RSNA Dataset ... 75
4.3 Simulation on Chest X-Ray Image Dataset ... 82
4.4 Findings ... 90
4.5 Summary ... 92
CHAPTER FIVE: CONCLUSION AND RECOMMENDATIONS ... 93
5.1 Conclusion ... 93
5.2 Future Scope ... 94
PUBLICATION ... 95
REFERENCES ... 96
APPENDIX: CODES ... 107
ix
LIST OF TABLES
Table 2.1 Suitable output activations and cost functions for different types
of classification problems 28
Table 2.2 Deep Learning techniques for Breast Image Analysis 41 Table 2.3 Deep Learning techniques for cardiac image analysis 42 Table 2.4 Deep Learning techniques for abdominal image analysis 43 Table 2.5 Overview of Deep Learning models for pneumonia detection from
chest x-ray Image Analysis 46
Table 3.1 Hyper- parameters of CNN training for RSNA dataset 71 Table 3.2 Hyper- parameters of CNN training for CXI dataset 71 Table 4.1 Analysis of variation of training and test sets over accuracy and
training time 88
Table 4.2 A comparison table for accuracy among different research works 88
x
LIST OF FIGURES
Figure 2.1 Biological neurons vs Artificial Neural Networks 11 Figure 2.2 Sample structure of a feed forward neural structure with three
hidden layers 12
Figure 2.3 The Supervised Learning model 13
Figure 2.4 The Unsupervised Learning model 14
Figure 2.5 The Reinforcement Learning model 15
Figure 2.6 An illustration of the different layers in a convolutional neural
network 16
Figure 2.7 Three examples of simple two-dimensional 3×3 edge detecting
filters 17
Figure 2.8 Convolution of an image and a small filter 19
Figure 2.9 Connections between neurons in a fully connected layer (left) and
a sparsely connected layer (right). 21
Figure 2.10 Example of max pooling. 22
Figure 2.11 Different kinds of classification problems. The output of the
network is either 1 (positive) or 0 (negative) for each class. 24 Figure 2.12 Two networks with different output activation functions 26 Figure 2.13 Example plot of the predicted probability against the cross-
entropy loss 28
Figure 2.14 Training error and validation error over capacity. 30 Figure 2.15 A confusion matrix showing the four possible outcomes of
classification for some input. 32
Figure 2.16 An illustration of difference between “traditional” machine
learning and deep learning. 37
Figure 2.17 Chest x-ray with air-space consolidation 51
Figure 2.18 Chest x-ray with nodular consolidation. 52
Figure 2.19 Chest x-ray with interstitial consolidation. 54
Figure 2.20 Chest x-ray showing patchy consolidation 55
xi
Figure 3.1 General flowchart of the implementation 58
Figure 3.2 Flow chart with the actions applied in pre-processing 61 Figure 3.3 Dataset part of a sample DICOM file module. 64
Figure 3.4 Output of black border remover algorithm 65
Figure 3.5 Image showing the ResNet model 67
Figure 3.6 Image showing the removal of last layer for training 68 Figure 3.7 Image showing nodes and addition of new prediction layer 68 Figure 3.8 Image showing the addition of new prediction layer 69 Figure 3.9 A common block in a regular stacked network (left) and a
residual block in the ResNet architecture (right). 70
Figure 3.10 Accuracy and Loss plot for RSNA dataset 72
Figure 3.11 Accuracy and Loss plots for CXI dataset 73
Figure 4.1 (a) Accuracy versus Epoch plot 76
Figure 4.1 (b) Loss versus Epoch plot 76
Figure 4.2 (a) Accuracy versus Epoch plot 77
Figure 4.2 (b) Loss versus Epoch plot 77
Figure 4.3 (a) Accuracy versus Epoch plot 78
Figure 4.3 (b) Loss versus Epoch plot 79
Figure 4.4 (a) Accuracy versus Epoch plot 79
Figure 4.4 (b) Loss versus Epoch plot 80
Figure 4.5 (a) Accuracy versus Epoch plot 81
Figure 4.5 (b) Loss versus Epoch plot 81
Figure 4.6 (a) Accuracy versus Epoch plot 82
Figure 4.6 (b) Loss versus Epoch plot 82
Figure 4.7 (a) Accuracy versus Epoch plot 83
Figure 4.7 (b) Loss versus Epoch plot 83
Figure 4.8 (a) Accuracy versus Epoch plot 84
xii
Figure 4.8 (b) Loss versus Epoch plot 85
Figure 4.9 (a) Accuracy versus Epoch plot 86
Figure 4.9 (b) Loss versus Epoch plot 86
Figure 4.10 (a) Accuracy versus Epoch plot 87
Figure 4.10 (b) Loss versus Epoch plot 87
xiii
LIST OF SYMBOLS
x Input data y Output data
h(x) Hypothesis function
θi Weight associated with neuron
b Bias
α Activation function i ith layer
j jth layer
Zj Output activation for the jth layer
* Convolution 𝜏 Dummy variable
I Input image K Filter matrix S Feature map
xiv
LIST OF ABBREVIATIONS
AE Auto Encoders
AI Artificial Intelligence
AIDS Acquired Immune Deficiency Syndrome ANN Artificial Neural Network
API Application Programming Interface
AUC Area Under Curve
CNN Convolutional Neural Networks CPU Central Processing Unit
CT Computerized Tomography
CXR Chest x-ray
CXI Chest x-ray Image Dataset
DICOM Digital Imaging and Communication on Medicine
DL Deep Learning
DNN Deep Neural Networks
DRL Deep Reinforcement Learning
FN False Negative
FP False Positive
GAN General Adversarial Network GPU Graphics Processing Unit
ID Identity Document
ILSVRC ImageNet Large Scale Visual Recognition Challenge JPEG Joint Photographic Experts Group
LSTM Long Short Term Memory
MG Mammography
ML Machine Learning
MLP Multilayer Perceptron
MRI Magnetic Resonance Imaging NIH National Institute of Health PCA Principle Component Analysis ResNet Residual Network
ROC Receiver Operating Characteristics RSNA Radiological Society of North America SGD Stochastic Gradient Descent
TB Tuberculosis
TDL Temporal Difference Learning
TL Transfer Learning
TN True Negative
TP True Positive
1
CHAPTER ONE INTRODUCTION
1.1 INTRODUCTION
Pneumonia remains the leading killer of young children in spite of the disposal of simple, safe, effective and economical treatment to reduce its capacity to kill (Wardlaw et al., 2006). In most of the articles and conferences it is known as "the silent killer", a nickname that shows little political and social awareness as compared to other diseases and its number of affectations is on increase each year. Childhood pneumonia is mainly a disease of poverty and is caused mostly because of inappropriate childcare and education on top of the unavailability of the access to healthcare. Adult pneumonia can be considered as a public health problem and needs more active involvement to cure it.
In South East Asian (SEA) Region, the rate of occurrence of pneumonia in children under five years of age is estimated as 0.36 episodes per child year as compared to the world average which is 0.26. In the context of developing countries average is found to be 0.29 while for developed countries is 0.05 episodes per child year.
61 million yearly new cases of childhood pneumonia is found in SEA Region out of a total of 156 million worldwide. 19 percent deaths among the under-five population in SEA Region is estimated to occur because of pneumonia as compared to 3.1 million annual deaths. Importantly these deaths do not include pneumonia cases among neonatal infections (Rudan et al., 2008).
In addition to the leading cause of death among children, lower respiratory infections are also the leading causes of disability-adjusted life years (DALYs) in SEA region as well as across the world. Almost 30 million DALYs and 1.4 million deaths among all age groups in SEA regions is found every year. This number exclude DALYs
2
and deaths caused by other respiratory diseases such as measles, whooping cough or TB (Mathers & Organization, 2008).
Accurate diagnosis of pneumonia is a tall order. It needs review of chest radiograph (CXR) by highly trained radiologists and confirmation through clinical history, vital signs and laboratory exams. Pneumonia being a disease of infectious origin is manifested as an area or areas of increased opacity on CXR (Franquet, 2018). The immune response of our body against this infection causes the lungs to fill with pus or other liquids which reduces its ability to hold air, and hence the patient may feel a choking sensation, cough and fever among other symptoms. However, because of a number of other conditions in the lungs such as volume loss (atelectasis or collapse), bleeding, fluid overload (pulmonary edema), lung cancer or post-radiation or surgical changes, the diagnosis of pneumonia in CXR becomes very complicated. Outside of the lungs, fluid in the pleural space (pleural effusion) also appears as increased opacity on CXR. A pneumonia opacity is a part of the lungs that looks darker on a radiograph and has a shape that indicates that pneumonia is (or may be) present.
If available, comparison of CXRs of the patient at different points of time and its correlation with clinical history and symptoms are helpful in the diagnosis of pneumonia.
Chest radiography being inexpensive and easy-to-use is the most common medical imaging and diagnostic technique. Even in underdeveloped areas, modern digital radiography (DR) machines are available at reasonable costs. Therefore, CXRs are widely used in the detection and diagnosis of lung diseases such as tuberculosis, pulmonary nodules, pneumonia, early lung cancer etc.
A large amount of information about the patient’s health is contained in the CXRs, but interpreting the information correctly is always a big challenge for the doctors and radiologists. Complexity of interpretation increases with the overlapping of
3
the tissue structures in the chest x-ray. For instance, when the contrast between the lesion and the surrounding tissue is very low or when the lesion overlaps the ribs or large pulmonary blood vessels, detection becomes challenging. Even if the doctor is experienced, sometimes it is not easy to differentiate between similar lesions or to find very obscure nodules. Consequently, the examination of disease in the chest x-ray will result in a certain degree of missed detection.
This inconsistency in the interpretation of medical images gives rise to a paradigm shift from a completely manual system of interpretation by professionals to the introduction of CAD (Computer-Aided-Diagnosis) systems to help these professionals in better interpretation and diagnosis while minimizing human errors.
Also, these systems will help in training novice technicians having less experience and standardize the reading and interpretation of CXRs.
Artificial Intelligence methods (including shallow learning and deep learning, etc.), mostly deep learning, replaces the step of feature extraction and disease classification in the traditional CAD systems. Artificial Intelligence has also played a big role in image segmentation and bone suppression of chest x-ray. The shallow learning methods are commonly used as classifiers in the detection of diseases, but their performance mostly depends on the extracted handcrafted features. If the chest x-ray images are complex, they take a lot of time and effort for researchers to find a good set of features that will be helpful for a good CAD performance. But now we are living in the age of Deep Learning algorithms and more specifically Convolutional Neural Networks (CNN) which are able to learn and extract features and able to represent very complex functions by themselves based solely on input data after a supervised training process without human interference. The ability of neural networks to learn by itself has
4
opened up promising prospects for their application in the interpretation and analysis of radiographic images.
Recently, due to the extensive growth and achievements of deep learning in various fields like image recognition (such as image classification (He et al., 2016;
Krizhevsky et al., 2012) and semantic segmentation (Long et al., 2015; Mostajabi et al., 2015), interest has been stimulated in reapplying deep learning to medical images.
Numerous studies and works confirm how Deep Learning algorithms have achieved a performance at par or superior to humans in various tasks and applications. A very famous and well known case was documented on the ImageNet Large-Scale Visual Recognition Challenge (ILSVRC) when a model overtook the human-level performance in the task of classifying images in 2015 (He et al., 2015). Starting from there, the models and algorithms based on Deep Learning continued to improve in the field of computer vision not only in the task of image classification but also in various other tasks such as object detection or segmentation, just to name a few.
In particular, the development of increasingly powerful GPUs coupled with the availability of large datasets has made these resource intensive models possible and have made these models perform better than the medical professionals in a wide variety of medical imaging tasks, including pneumonia diagnosis (Rajpurkar, Irvin, et al., 2017), diabetic retinopathy detection (Gulshan et al., 2016), skin cancer classification (Esteva et al., 2017), arrhythmia detection (Rajpurkar, Hannun, et al., 2017), and bleeding identification (Grewal et al., 2018). Therefore, Deep Learning methods (especially CNNs), which automatically learn image features to classify chest diseases, have become a mainstream trend.
Now focusing on the area concerned i.e. the diagnosis of pneumonia cases from chest x-rays it has been found that only a dataset called "Chest x-ray 14" created by the
5
NIH (National Institute of Health) is available with annotations and it is focused mainly on the tasks of classification of pulmonary pathologies. The availability of the above dataset resulted in various works focused on the classification of pulmonary pathologies based on chest x-rays (Rajpurkar, Irvin, et al., 2017; Wang et al., 2017; Yao et al., 2018).
But due to the absence of a set of data with necessary annotations the task of detection and location of anomalies was not possible to develop.
Radiological Society of North America (RSNA) has recently published a dataset which is focused mainly on object detection and classification cases of pneumonia. The publication of this new dataset opens up the possibility for the development of more advanced Deep Learning models for the detection and diagnosis of pneumonia. Using the current dataset, there is possibility of improving the existing classification models and at the same time developing new object detection architectures in order to make the system more competent in the detection and localization of pneumonia cases from radiographic images.
1.2 PROBLEM STATEMENT
Pneumonia is a major killer of old age people and children under five years of age all around the world. It is an infection caused by bacteria, virus or other germs and results in the swelling of lungs which can be life threatening if not diagnosed in time.
Chest x-ray is the most common diagnostic tool used in medical practice.
Pneumonia is manifested as an area or areas of increased opacity in the CXR. However, correct interpretation of chest x-rays for the diagnosis of pneumonia is always a challenge for the doctors and radiologists. This is mainly because several other medical conditions such as lung cancer, excess fluid etc. can also show similar opacities in images. Also, the complexity of interpretation increases with the overlapping of tissue
6
structures. All these reasons make the interpretation of pneumonia from chest x-rays complex and less accurate. Therefore, there is a need for a system for accurate interpretation of chest x-ray images.
Although several architectures were used before for the classification of disease from chest x-rays but none of them gave a satisfactory result. For instance, VGG16 is a comparatively smaller network. It may not be able to find the necessary features needed for pathology classification. That could be the reason for its poor performance. Training a full Densenet-121 network took longer time to train but gave similar results as compared to VGG16. This could be a consequence of using such a deep network architecture. When too many layers are added to a model, accuracy gets saturated and the model starts degrading. In order to overcome the shortcomings stated above, this project aims to develop a model for effective detection of pneumonia from chest x-rays.
1.3 RESEARCH OBJECTIVES
The main goal of the research is to design a model based on Deep Learning that can classify chest x-rays into cases of pneumonia and non-pneumonia. Two datasets namely RSNA Pneumonia Detection Challenge dataset provided by Radiological Society of North America and chest x-ray Image dataset provided by Kermany et al., will be used.
These datasets provide the necessary annotations needed to train and evaluate the classification model. The specific objectives can be stated as following:
• To develop a classification model based on transfer learning technique that can classify chest x-rays into pneumonia and non-pneumonia images using Deep Convolutional Neural Networks.
• To evaluate the developed model with the test dataset.
7
• To benchmark the developed model’s accuracy with the relevant research works.
1.4 RESEARCH METHODOLOGY
The main objective of the research is to develop an effective model for the classification of chest x-rays into pneumonia and non-pneumonia cases. The specific methodology followed to achieve the tasks are as follows:
i. Dataset: Identification of two datasets which provides a wide variety of images of pneumonia and are mostly used among the researchers in the common field.
ii. Pre-processing: In this step we will focus on the actions applied on the dataset samples as a previous step to feed the models of the classification during training and prediction process. Adding, removing and transforming attributes are the three types of pre-processing which are considered for the data. Also, the dataset is split into training, validation and test sets for classification and validation task.
iii. Transfer Learning: In this step the model is trained. For training the model transfer learning technique is used.
iv. Classification: The Resnet-50 model is used for classification task.
v. Prediction: In this step a random chest x-ray image is loaded in the prediction model and is checked for pneumonia. A prediction accuracy is given as output.
8 1.5 RESEARCH SCOPE
Primary objective of the research is to develop a model for classification of pneumonia.
The current work is focussed mainly on the classification task. Detection and localization of pneumonia in the radiographic image has not been considered. Also, the research will be conducted on two datasets which are especially selected for this task.
Although 14 types of diseases can be detected and diagnosed through chest x-ray, but the current work concentrates only on the detection of pneumonia.
1.6 THESIS ORGANIZATION
The current research work has been documented in five chapters as follows:
• Chapter one gives an introduction of the topic along with the context.
Research problem is discussed along with specific objectives. And finally, methodology to be followed and research scopes are discussed.
• Chapter two gives the theoretical background needed to understand the deep learning concepts. In this chapter, first of all, the reader is provided with a general literature review covering the use of deep neural networks in the diagnosis of diseases through chest radiographic images. After that a critical review is done in the concerned area i.e. detection of pneumonia through chest x-rays. And finally, interpretation cases of pneumonia in chest x-rays is presented. Within this various type of patterns of consolidation on chest x-rays concerning pneumonia are discussed.
• Chapter three discusses the research methodology. In this chapter first of all a general structure of the proposed implementation is given. After that a discussion on framework which is used for the implementation of the model is discussed. Finally, the methodology to be followed is discussed in detail.
9
• Chapter four discusses about the various simulations done. Later on, the results are presented along with the findings.
• Chapter five is dedicated to the conclusions drawn from the project and the introduction to the future recommendations.