Classification of Lung Diseases Using Deep Learning Models

Classification of Lung Diseases Using Deep Learning
Models
Matthew Zak
A Thesis
in
The Department
of
Computer Science and Software Engineering
Presented in Partial Fulfillment of the Requirements
for the Degree of
Master of Computer Science at
Concordia University
Montréal, Québec, Canada
September 2019
c Matthew Zak, 2019

C ONCORDIA U NIVERSITY
School of Graduate Studies
This is to certify that the thesis prepared
By: Matthew Zak
Entitled: Classification of Lung Diseases Using Deep Learning Models
and submitted in partial fulfillment of the requirements for the degree of
Master of Computer Science
complies with the regulations of this University and meets the accepted standards with respect to
originality and quality.
Signed by the Final Examining Committee:
Chair
Dr. Andrew Delong
External Examiner
Dr. Thomas Fevens
Examiner
Dr. Ching Yee Suen
Supervisor
Dr. Adam Krzyzak
Approved by
Lata Narayanan, Chair
Department of Computer Science and Software Engineering
2019
Amir Asif, Dean
Faculty of Engineering and Computer Science
iii
Abstract
Master of Computer Science
Classification of lung diseases using deep learning models
by Matthew Z AK
Although deep learning-based models show high performance in the medi-
cal field, they required large volumes of data which is problematic due to pro-
tection of patient privacy and lack of publically available medical databases.
In this thesis, we address the problem of medical data scarcity by consid-
ering the task of pulmonary disease detection in chest X-Ray images using
small volume datasets (<103 samples). We implement three deep convolu-

tion neural networks pre-trained on the ImageNet dataset (VGG16, ResNet-
50, and InveptionV3) and asses them in the lung disease classification tasks
transfer learning approach. We created a pipeline that applied segmentation

on Chest X-Ray images before classifying them and we compared the per-
formance of our framework with the existing one. We demonstrated that
pre-trained models and simple classifiers such as shallow neural networks

can compete with the complex systems.
We also implemented activation maps for our system. The analysis of class
activation maps shows that not only does the segmentation improve results
in terms of accuracy but also focuses models on medically relevant areas of
lungs.
iv
We validated our techniques on the publicly available Shenzhen and Mont-
gomery datasets and compared them to the currently available solutions.

Our method was able to reach the same level of accuracy as the best per-
forming models trained on the Montgomery dataset however, the advantage

of our approach is a smaller number of trainable parameters. What is more,
our InceptionV3 based model almost tied with the best performing solution
on the Shenzhen dataset but as previously, it is computationally less expen-

sive.
v
“Everything has been, everything has happened. And everything has already been
written about.”
Vysogota de Corvo
“There is nothing noble in being superior to your fellow man; true nobility is being
superior to your former self.”
Ernest Hemingway
vi
Acknowledgements
I want to thank everyone who has had the tiniest input in this thesis. If it had
not been for you, I would probably have never made it.
vii
Contents
Declaration of Authorship ii
Abstract iii
Acknowledgements v
1 Introduction 1
1.1 Project objectives . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.2 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.4 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.5 Thesis structure . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2 Related Work 4
2.1 Pixel/Voxel-Based Machine Learning . . . . . . . . . . . . . . 4

2.1.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1.2 Bone Separation from Soft Tissue in Chest Radiographs

(CXRs) by Use of MTANNs . . . . . . . . . . . . . . . . 6
2.2 Approaches to lung disease classification . . . . . . . . . . . . 8
2.2.1 Extreme learning machine . . . . . . . . . . . . . . . . 8
2.2.2 Automatic CXR screening system . . . . . . . . . . . . 10

2.2.3 Semi-Supervised Learning . . . . . . . . . . . . . . . . . 12
2.3 Automatic diagnostic using Deep Learning in MODS digital
images analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.4 Deep Learning Approaches in Chest X-Ray analysis . . . . . . 17
viii
2.5 Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.6 Image Data Augmentation . . . . . . . . . . . . . . . . . . . . . 20
2.7 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.8 Convolutional Neural Networks . . . . . . . . . . . . . . . . . 26
2.9 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3 Transfer Learning in Lung Diseases Classification 30

3.1 Transfer learning . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.1.1 Pre-trained models approach . . . . . . . . . . . . . . . 31
3.1.2 ImageNet . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.2 VGG16 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.2.1 VGG16 Architecture . . . . . . . . . . . . . . . . . . . . 32

3.3 ResNet-50 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.3.1 Residual Bloc . . . . . . . . . . . . . . . . . . . . . . . . 35
3.3.2 ResNet50 Architecture . . . . . . . . . . . . . . . . . . . 36
3.4 Inception . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.4.1 Factorizing Convolutions . . . . . . . . . . . . . . . . . 40

3.4.2 Factorization Into Asymmetric Convolutions . . . . . . 40
3.4.3 Auxiliary Classifiers . . . . . . . . . . . . . . . . . . . . 42
3.4.4 Effective Grid Size Reduction . . . . . . . . . . . . . . . 43
3.4.5 Architecture . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.5 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.5.1 Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.5.2 Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.6 Results and analysis . . . . . . . . . . . . . . . . . . . . . . . . 47
3.6.1 VGG results . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.6.2 ResNet results . . . . . . . . . . . . . . . . . . . . . . . . 53

3.6.3 Inception results . . . . . . . . . . . . . . . . . . . . . . 58
3.6.4 Results comparison . . . . . . . . . . . . . . . . . . . . . 63

ix
3.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4 Transfer Learning Models Accuracy Improvement in Lung Diseases
Classification Using Segmentated X-Ray Images 66

4.1 U-Net - Image Segmentation using Deep Neural Networks . . 66
4.1.1 Architecture . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.2 Lungs Segmentation . . . . . . . . . . . . . . . . . . . . . . . . 69

4.2.1 Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.2.2 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.2.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.3 Training Deep Learning Models On Segmented Images . . . . 73
4.3.1 Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.4 Results and analysis . . . . . . . . . . . . . . . . . . . . . . . . 73
4.4.1 VGG results . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.4.2 ResNet results . . . . . . . . . . . . . . . . . . . . . . . . 79
4.4.3 Inception results . . . . . . . . . . . . . . . . . . . . . . 84
4.5 Comparison of results . . . . . . . . . . . . . . . . . . . . . . . 90

4.6 Comparison with other works . . . . . . . . . . . . . . . . . . . 92
5 Conclusions and Future Work 94

5.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
5.2 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
5.3 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
Bibliography 97
x
List of Figures
2.1 Ribs separation training samples. Image A is the input and B

the corresponding boneless output. . . . . . . . . . . . . . . . . 7
2.2 Performance of a trained model in ribs removal task. Image A

is an input sample and B is a result with removed ribs . . . . . 8
2.3 Overview of the screening system developed for tuberculosis
detection [1] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.4 Overview of the CST-Voting algorithm [2] . . . . . . . . . . . . 14
2.5 CST-Voting algorithm [2] . . . . . . . . . . . . . . . . . . . . . . 14

2.6 The left image shows a tuberculosis cord in a white highlighted
box. The right part is a positive tuberculosis culture [3] . . . . 16

2.7 "Simplified network architecture. (A) Input to the network is a
224 x 224 grayscale image of a MODSM. tuberculosis culture.
The image is passed through the network, and the output of
the second fully-connected layer is a probability distribution

over the two classes (positive (+): 1 and negative (–): 0). Each
block is a stack of feature maps, of dimensions (width x height
x number of feature maps). Layer operations take place be-

tween each block (see (B)) and are identifiable by the feature
map volume produced. Kernels are 3 x 3 and 2 x 2 for convolu-
tional and pooling layers, respectively. The network is trained

and evaluated on a dataset of 1008 train/validation and 2502
test images. (B) A schematic representation of the convolution
and pooling operations on an input volume." [3] . . . . . . . . 16

xi
2.8 Sample Chest X-Ray images containing marks of tuberculosis

(A) and pneumonia (B). . . . . . . . . . . . . . . . . . . . . . . 18
2.9 "Combined distribution of genders and ages among images

without (0) and with (1) disease marks." [4] . . . . . . . . . . . 19
2.10 Example of an original image in SH dataset (left) and the seg-

mented result (right). . . . . . . . . . . . . . . . . . . . . . . . . 20
2.11 Distributions of image (Fig. 2.10, left) and mask (Fig. 2.10,
central) areas. [4] . . . . . . . . . . . . . . . . . . . . . . . . . . 20

2.12 Selected image transformation methods. Image A) shows the
original image, B) flipped, C) darkened (brightness change)

and finally D) zoomed(cropped and upsized) . . . . . . . . . . 21
2.13 Tp vs. FP using different thresholds of classification.[5] . . . . 25
2.14 AUC (Area under the ROC).[5] . . . . . . . . . . . . . . . . . . 26
2.15 Convolutional Neural Network for image processing.[6] . . . 27

2.16 A hypothetical 4x4 image. . . . . . . . . . . . . . . . . . . . . . 28
2.17 Convolution operation. . . . . . . . . . . . . . . . . . . . . . . . 28
2.18 Max pooling operation. . . . . . . . . . . . . . . . . . . . . . . . 29
3.1 VGG network with 16 layers. . . . . . . . . . . . . . . . . . . . 34
3.2 Residual bloc. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

3.3 Residual blocs . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.4 Resnet-34 architecture . . . . . . . . . . . . . . . . . . . . . . . 38

3.5 Inception cell introduced with the fist Inception model[7] . . . 39
3.6 5x5 convolution replaced by two 3x3 convolutions. . . . . . . 40
3.7 Inception cell introduced in [8] . . . . . . . . . . . . . . . . . . 41

3.8 Asymmetric convolution [8] . . . . . . . . . . . . . . . . . . . . 41
3.9 Inception cell using aymmetric convolutions[8] . . . . . . . . . 42
3.10 Wider inception module[8] . . . . . . . . . . . . . . . . . . . . . 43

xii
3.11 "Auxiliary classifier on top of the last 17x17 layer. Batch nor-
malization of the layers in the side head results in a 0.4% abso-
lute gain in top-1 accuracy. The lower axis shows the number
of iterations performed, each with batch size 32." [8] . . . . . . 43
3.12 "Two alternative ways of reducing the grid size. The solution
on the left violates principle 1 of not introducing a representa-
tional bottleneck from Section 2. The version on the right is 3
times more expensive computationally." [8] . . . . . . . . . . . 44

3.13 "Inception module that reduces the grid-size while expands
the filter banks. It is both cheap and avoids the representa-

tional bottleneck as is suggested by principle 1. The diagram
on the right represents the same solution but from the perspec-
tive of grid sizes rather than the operations." [8]. . . . . . . . . 45
3.14 InceptionV3 architecture. Batch normalization and ReLU non-

linearity are used after every convolution layer.) . . . . . . . . 45
3.15 VGG16 based model training and validation loss change. . . . 49
3.16 VGG16 based model training and validation accuracy change. 50
3.17 Image A shows the confusion matrix of the obtained results
with the model which reached the highest accuracy during the
training. Image B show its per class ROC curves. . . . . . . . . 52
3.18 Two pairs of correctly labelled images containing marks of tu-
berculosis and pneumonia with their class activation maps. . . 53

3.19 ResNet-50 based model training and validation loss change. . 55
3.20 ResNet-50 based model training and validation accuracy change. 56


xiii
3.23 InceptionV3 based model training and validation loss change. 60

3.24 InceptionV3 based model training and validation accuracy change. 61


4.1 "U-net architecture (example for 32x32 pixels in the lowest res-
olution). Each blue box corresponds to a multi-channel fea-
ture map. The number of channels is denoted on top of the
box. The x-y-size is provided at the lower left edge of the box.
White boxes represent copied feature maps. The arrows de-

note the different operations."[9] . . . . . . . . . . . . . . . . . 67
4.2 An X-Ray image and its corresponding lungs mask. . . . . . . 70
4.3 U-Net training and validation losses change during training. . 71
4.4 Three pairs of CXR images with corresponding, segmented
(extracted) lungs. . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.5 VGG16 based model training and validation loss change. . . . 75
4.6 VGG16 based model training and validation accuracy change. 76


4.9 ResNet-50 based model training and validation loss change. . 81

4.10 ResNet-50 based model training and validation accuracy change. 82
xiv



4.13 InceptionV3 based model training and validation loss change. 86
4.14 InceptionV3 based model training and validation accuracy change. 87


4.17 Two pairs of correctly classified images with their class activa-
tion maps. The left columns is non segmented images and the
right is the segmented ones. . . . . . . . . . . . . . . . . . . . . 91
xv
List of Tables
2.1 Different classes of PMLs, their functions and applications [10] 5

2.2 Dataset class distribution. . . . . . . . . . . . . . . . . . . . . . 19
2.3 Hypotetical data distribution . . . . . . . . . . . . . . . . . . . 22
3.1 Comparison of transfer-learning based algorithms in terms of
accuracy, AUC, F1 score, precision and sensitivity for healthy,
pneumonia and tuberculosis classes. . . . . . . . . . . . . . . . 64
4.1 Comparison of all results received on segmented and non-segmented
data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
4.2 Comparison of different deep learning based solutions trained
on the Shenzhen datases. Although our result is not the best, it
performs better than any single model (excluding Ensemble).
Horizontal line means that those results were not provided in
literature. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.3 Comparison of different deep learning based solutions trained
on the Montgomery dataset [11]. Our average performance is
almost identical to [12]. . . . . . . . . . . . . . . . . . . . . . . . 93

1
Chapter 1
Introduction
1.1 Project objectives
Computer vision supported with deep neural networks finds its usefulness
in any area of life starting from facial emotion recognition to disease detec-
tion. Thanks to the recent technological advances, computer-aided image
analysis algorithms compete with professionals in terms of accuracy yet re-

main unchallenged in speed and volume of reviewed cases. Unlike doctors,
computers make quick rational decisions unaffected by emotions and tired-
ness. The newest WHO (World Health Organisation) report states that just
in the United States over 1 million citizens seek care due to pneumonia and
there are nearly ten million cases of tuberculosis worldwide. Perhaps some
part of it is lethal due to lack of medical staff or human mistake. The man-
ual analysis of x-ray images is a long process requiring radiological expertise

and a large volume of time. Deep learning can play a crucial role in exceed-
ing decision making, detecting marks of disease as well as conducting the
initial examination and suggesting urgent cases.

Chapter 1. Introduction 2
1.2 Motivation
Previous approaches required both a large volume of data and strong com-
puting power computers [4]. Instead of following the popular trend of creat-
ing new algorithms to solve a problem, we decided to leverage existing tools

and show their high accuracy in medical tasks outperforming the suggested
solutions [13]. The motivation of this project was to create a pipeline allowing
us to detect pulmonary diseases using tiny datasets (< 103 images per class)
(see Chapter 2) and limited computational resources. Also, we want to show
the importance of segmentation(see Chapter 2 for details) as a tool focusing

our algorithms on information that is crucial in diagnosis. Should models
search through pulmonary changes within lungs images omitting redundant
data (bones, internal organs)[14], the decision is considered reasonable. To
prove that statement, we generate class activation maps marking regions of
interest (regions that vastly contributed to the final classification).
1.3 Applications
As mentioned before, the objective of this work is to create a model able
to provide relatively good results even if the resources are limited (both data
and computational). We believe that hospitals could find this solution useful,
especially due to data confidentiality and graphical processing unit (GPU)

accessibility.
1.4 Contributions
This work brings a new outlook on pre-trained deep neural networks and
their combinations with segmentation models. It makes the following con-

tributions, both theoretical and practical:
Chapter 1. Introduction 3
1. Segmentation as a mean of focusing transfer learning-based neural net-

works on regions of interest (containing marks of lung diseases).
2. Transfer learning based classifiers trained on small Chest X-Ray data
sets [4] (see chapter 3) and the positive impact of segmentation on la-
beling accuracy (see chapter 4)
3. Application of different deep learning models in lung diseases problem

(see chapter 3 and 4)
4. The usefulness of transfer learning in deep networks training on a small
medical dataset (see chapter 3)
1.5 Thesis structure
The first chapter serves as an introduction and motivation to the topic of lung
disease classification. We briefly described the contributions and meaning-
fulness of a given task, yet the real reason is about making the world a better
place.
The second chapter describes the essential background and significant re-
lated work done in lung disease classification. We present neural networks
as a solution to the problem, datasets, and results in evaluation metrics.

4
Chapter 2
Related Work
As this dissertation focuses on lung diseases (pneumonia and tuberculosis)
classification, we first want to discuss selected screening approaches briefly.
All techniques use radiography, a medical imaging method that used to be

obsolete. However, machine learning and digital advances revived this method
[15] and its significance in diagnosis of lungs diseases [16][17][18][19]. Espe-
cially, they allow detecting multiple forms of cardiothoracic lessons on x-ray
scans. The growing popularity of machine learning models in medical diag-
nosis correlated with their accuracy is considered as a great success as it leads

to better disorder recognition. Recent encouraging results in deep learning
applied in the field of lung diagnosis led to the usage of a GPU-based plat-
form which is able to process a large volume of images in high-resolution
within seconds and thus exceed the work of radiologists.
2.1 Pixel/Voxel-Based Machine Learning
The availability of computationally powerful machines allowed emerging
methods like pixel/voxel-based ML (PML) in medical image analysis/processing.

Instead of calculating features from segmented regions, this technique uses
voxel/pixel values in input images directly. Therefore, neither segmenta-

tion nor feature extraction is required. The performance of PML can possibly
Chapter 2. Related Work 5
get higher than common classifiers[10] as this method is able to avoid errors
caused by inaccurate segmentation and feature extraction.
2.1.1 Overview
Medical image processing/analysis, as well as computer vision, have adopted

many PMLs in many tasks. The Table 2.1 provides a brief summary of PML
classes, their functions and applications.
PMLs Functions Applications

Enhancement of subjective edges
traced by a physician [20].
Neural filters
Edge enhancement from
(including neural Image processing
noisy images [21].
edge enhancers)
Edge-preserving noise
reduction [22][23].
FP reduction in CAD for detection
of masses [24]
Convolution neural and microcalcifications [25] in
networks (including mammography.
Classification
shift-invariant neural FP reduction in CAD for lung
networks) nodule detection in CXR [26][27].
Character recognition [28].
Face recognition [29].
FP reduction in CAD for detection
of lung nodules in
Massive-training
Classification CXR and CT [30].
artificial neural
(image processing Distinction between benign
networks (MTANNs,
+ scoring), object and malignant lung nodules in CT.
including a mixture
detection, suppression FP reduction in CAD for polyp
of expert MTANNs,
and pattern detection in CT colonography.
a LAP-MTANN,
enhancement Bone separation from soft tissue
an MTSVR)
in CXR. Enhancement of lung
nodules in CT.
Separation of ribs from soft tissue
Image processing
Others in CXR [31]. Segmenting posterior
or classification
ribs in CXR [32].
TABLE 2.1: Different classes of PMLs, their functions and ap-

plications [10]
We can distinguish three classes of PMLs [10]: convolution neural net-

works [27] [33][24] (this includes shift-invariant Neural Networks [34]), neu-
ral filters [23] and MTANNs or massive-training artificial neural networks

[35] (including multiple variations, i.e., Laplacian eigenfunction MTANN or
massive-training support vector regression). Many image analysis tasks used

neural filters. Some of them are edge enhancement from noisy images, edge-
preserving noise reduction in digital pictures and radiographs [23], and en-
hancement of subjective edges traced by a physician in cardiac ventriculo-
grams. Convolutional neural networks have been applied to classification
in tasks requiring false-positives reduction in CAD schemes for detection of

lung nodules [27][26][33] or microcalcifications in chest X-rays, masses in
mammography, character, and face recognition. Massive-training artificial
neural networks have been used in classification problems such as a false-
positive reduction in CAD schemes for detection of lung nodules [30], the
distinction between malignant nodules and the benign ones. This class of
PMLs has also been applied to suppression, and pattern enhancement of
bone tissue from soft tissue in CXR [35] and lung nodules enhancement in
X-ray computed tomography.
2.1.2 Bone Separation from Soft Tissue in Chest Radiographs
(CXRs) by Use of MTANNs
Chest X-Ray is one of the most frequently used diagnosis when examining
medical images for different lung diseases such as pneumonia or tubercu-

losis. Roughly 1 million of adults require hospitalization because of pneu-
monia, and about 50,000 dies from this disease annually in the US only [36].
Examination of lung nodules in CSR can lead to overlooking of diseases like

lung cancer. However, not all of them are visible in retrospect. Studies show
that 82-95% of lung cancer cases were missed due to at least partially ob-
scured by ribs or clavicle. To address this issue, researchers examined dual-
energy imaging, a technique which can produce images of two tissues, which
is a "soft-tissue" image and a "bone" one [14]. This technique has many draw-
backs, but undoubtedly one of the most important ones is the exposure to
radiation.
The MTANNs models have been developed to address this problem and
serve as a technique for ribs/soft-tissue separation. The idea behind train-
ing those algorithms is to provide them with bone and soft-tissue images
obtained from a dual-energy radiography system. The MTANN was trained

using CXRs as input and corresponding boneless images, as presented in
Figure 2.1. Figure 2.2 shows the performance of the model on the unshown
image data. The ribs contrast is visibly suppressed in the resulting image,
maintaining the soft tissue areas such as lung vessels.
( A ) Origi- ( B ) Bone-
nal input less image
F IGURE 2.1: Ribs separation training samples. Image A is the

input and B the corresponding boneless output.
( A ) Test ( B ) Gen-
sample erated
result
F IGURE 2.2: Performance of a trained model in ribs removal

task. Image A is an input sample and B is a result with removed
ribs
2.2 Approaches to lung disease classification
2.2.1 Extreme learning machine
The grey-level co-occurrence matrix inspired the feature extraction method
used in [37], a technique often used in the texture analysis. The proposed
method named Spatial Interdependence Matrix or SIM makes use of the co-
occurrence statistics to analyze the structural information based on the way
human visual system tries interpreting scenes. In this context, the new tech-
nique can be implemented to evaluate image structural degradation.

Let us consider an image I and its degraded version J as one set of grey
levels in the domain D defined as:
D ⇢ Z2 2 0, 1, 2, 3, ..., N (2.1)
where N represents the number of all grey levels. The transition of pixel in-
tensities and spacial correspondence between images I and J is arranged into
a two dimensional matrix of size NxN, where each element Mi,j is defined as:
Mi,j = #(i, j) : i = I ( p) , j = J ( p) 8 p 2 D (2.2)
Where the #. operator represents the set cardinality of I(p) and J(p) intensities
for all pixels p, which belongs to both I and J.

The three structural attributes extracted from the Spatial Interdependence
Matrix matrix relate to the image J (degraded) when compared to I [37].

Those attributes are chi-square (Chi), inverse difference moment (Idm) and
correlation (Corr), and represent the degradation level in three different per-
spectives: structural independence, structural degradation, structural simi-
larity, respectively.
To extract the structural attributes Idm and Corr, the authors in [37] used
M+ MT
an asymmetric version of MS = 2 , where M T is simply the transposed
matrix M. The matrix M is normalized to obtain the pair-transitions weight
as a probabilities approximation. The structural attributes in terms of Mi,j

are described as follows:
N 1 (i µi ) j µ j Mi,j
Corr = Â q 2 [ 1, 1] (2.3)
i,j=0 si2 sj2
N 1 Mi,j
Idm = Â 1 + |i j |
2 [0, 1] (2.4)
i,j=0
N 1
(Oi Ei )2
Chi = Â Ei
(2.5)
i =0
N
i
µi = ÂN (2.6)
i =0
where µ j and µi correspond to the average values and sj and si relate to

the standard deviation of the matrix M for each column j and row i. Oi and Ei
stand the observed and expected weights respectively, in the diagonal matrix
of M (i = j) [37].
The Spatial Interdependence Matrix gives us a visual pattern which is

useful to interpret degraded images. If an image is undegraded, the weights
are distributed near the diagonal of the matrix. Otherwise, different patterns
appear depending on the degradation of the structure. When relating to the
lung’s diseases, the SIM pattern of fibrosis varies from one image to another
image of the healthy lungs. The structures of fibrosis are spread through the
lungs area, whereas in healthy lungs they are sparse and small.
A set of attributes extracted from the Spatial Interdependence Matrix (Cor-

relation, chi-square, and inverse difference moment) is used to asses the struc-
tural characterization of lung images. Using prior knowledge that the CT
images have blurred structures, authors in [37] convolve the inputs with a
gaussian kernel. They set the number of gray levels to 64 to compute the
Spatial Interdependence Matrix.
The presented lung disease descriptor [37] consists of three attributes in

a vector A = [Corr, Idm, 1 Chi ], which are extracted from the SIM matrices
of training images. Later on, they trained a multi-layer perceptron (MLP) for
100 epochs. The trained models allowed them to reach the accuracy level
above 96%
2.2.2 Automatic CXR screening system
This method is a multistage processing system based on a multistage frame-
work developed (see Figure 2.3) by researchers in [1] which first segment
images and then by combining texture and shape features tries to predict a
disease presented on CXRs. The algorithm uses similar intuition as radiol-
ogists during lung examination, which is the comparison of right and left
lung fiends. The texture features describe the inside lungs fields, and the
shape features focus on including the relevant geometrical characteristics
F IGURE 2.3: Overview of the screening system developed for

tuberculosis detection [1]
The presented method on 2.3 consists of three consecutive stages, which

are segmentation, feature extraction, and classification, respectively. The
first stage is composed of three phases. The first one is content-based im-
age retrieval using Bhattacharyya shape similarity measure [38] and a par-
tial Radon transform [39]. Then, building a patient anatomic model of lung
shape based on SIFT-flow [40]. And finally, taking out lungs boundaries with
an approach of graph cut optimization. The next stage is texture feature ex-
traction of segmented lungs using features such as intensity histogram, gra-
dient magnitude, a histogram of oriented gradients, etc. The final step is

classification using a support vector machine model (SVM) [41].
2.2.3 Semi-Supervised Learning
This section briefs the SSL algorithm proposed in [2] for the classification
of the pulmonary disease, which uses on the ensemble method called CST-
Voting [42]. The idea behind this algorithm is to generate a classifier by ap-
plying multiple Semi-Supervised Learning methods to one dataset. Using
this theory, the researchers in [2] created an ensemble model consisting of
three algorithms, which are: self-training, co-training, and tri-training.
Self-training algorithm implements a simple arbitrary model that is trained
on a small subset of labeled data L and tries to predict U. If the probability of

a predicted instance is higher than the defined confidence level, it is added to
L. We repeat this procedure until the set U is empty, or we meet some stop-
ping criterion.
Co-training assumes that we have weak algorithms that are trained on the
set of labeled instances (L). Later, the two algorithms classify instances of the
unlabeled set U’ (fixed-size subset of all unlabeled data U) and move to the L
those where the prediction was the most confident. After removing samples
from U’, we refill this set with new instances from U. We repeat this proce-
dure until the set U is empty, or we meet some stopping criterion.
Tri-training is an extension of the co-training method. Similarly to the previ-

ous approach, we have multiple weak learners, yet this time, its number is
increased to three. Each classifier is trained on the labeled data and predicts
the class for instances in the unlabeled set. The majority makes the final de-
cision, and the classified sample is added to the labeled data. One way of
looking at this method is that the majority teaches the minority - the majority
of votes decides about the final class for all learners.

Those self-labeled methods operating in different manners take full advan-
tage of the encoded information from the unlabelled data set. The crucial
feature of making those methods different is the mechanism behind labeling

the unlabelled data. Tri-training, as well as for self-training, are single-view
algorithms, whereas co-training is a multi-view one. What is more, both tri-
training and co-training are ensemble methods themselves. An overview
of the described CST-Voting algorithm is presented in Figure 2.4. At first,

all the mentioned semi-supervised algorithms (tri-training, co-training, and
self-training), which build the ensemble learn using the same unlabelled U
and label data set L. Afterwards, the decision on an unlabelled test sample
combines all separate predictions of the semi-supervised models. Moreover,
here, a simple voting methodology is applied - make a decision based on the
majority of votes [2]. The high-level explanation of the described CST-Voting

algorithm is shown in Figure 2.5.
F IGURE 2.4: Overview of the CST-Voting algorithm [2]
F IGURE 2.5: CST-Voting algorithm [2]

2.3 Automatic diagnostic using Deep Learning in
MODS digital images analysis
The Microscopic Observed Drug Susceptibility (MODS) is a test to analyze

tuberculosis contamination and drug resistance using a sputum test in about
one week period with minimal effort and high effectiveness. In spite of its
points of interest, MODS is as yet restricted in remote, low asset settings,
since it requires changeless and prepared specialized staff for the picture
based diagnostics. Henceforth, it is essential to create elective arrangements,
given solid mechanized investigation and elucidation of MODS societies. Re-
searchers in [3] trained and then assessed a deep convolutional neural net-
work 2.8 for MODS digital images interpretation and diagnostics.
The researchers used a dataset which consisted of almost 13 thousand

MODS digital images, as shown in Figure 2.6. It contained only two classes
od samples - positive, containing marks of tuberculosis and the negative
(healthy) ones. Similarly to this thesis, [3] used an adaptation of VGG16
network [43] (also see Section 3.2). A 15 layers deep model comprising fully-
connected, max-pooling, convolutional layers organized into five blocks; four
convolutional layers separated by pooling layers terminated by one fully-
connected layer (classifier). The architecture is given in Figure 2.7

F IGURE 2.6: The left image shows a tuberculosis cord in a white

highlighted box. The right part is a positive tuberculosis culture
[3]
F IGURE 2.7: "Simplified network architecture. (A) Input to the

network is a 224 x 224 grayscale image of a MODSM. tubercu-
losis culture. The image is passed through the network, and the
output of the second fully-connected layer is a probability dis-
tribution over the two classes (positive (+): 1 and negative (–):
0). Each block is a stack of feature maps, of dimensions (width x
height x number of feature maps). Layer operations take place
between each block (see (B)) and are identifiable by the feature
map volume produced. Kernels are 3 x 3 and 2 x 2 for convolu-
tional and pooling layers, respectively. The network is trained
and evaluated on a dataset of 1008 train/validation and 2502
test images. (B) A schematic representation of the convolution
and pooling operations on an input volume." [3]
2.4 Deep Learning Approaches in Chest X-Ray anal-
ysis
Recent applications of Deep Neural Networks [44][43][45][8][7] lead to major

improvements in medical imaging. The efficiency of dimensionality reduc-
tion algorithms like lung segmentation was demonstrated in the Chest X-Ray
image analysis [46]. Researchers in [4] aimed at improving tuberculosis de-
tection on relatively small data sets (< 103 images per class) by incorporat-
ing deep learning segmentation and classification methods from [46]. The
further exploration of these techniques is the topic of this thesis. Therefore

we focus on various deep learning methods in lung diseases classification in
chapters 3 and 4
2.5 Dataset
This work combines two relatively small datasets (< 103 images per class)
datasets for the classification (pneumonia and tuberculosis detection) and
segmentation purposes. We selected 306 examples per "diseased" class (306

images containing marks of tuberculosis and 306 images with pneumonia)
and 306 of healthy patients contributing to a set of 918 samples coming from
different patients. Sample images coming from both datasets are presented
in Figure 2.8.
The Shenzhen Hospital dataset (SH) [13][47] containing CXR images was
created by People’s Hospital in Shenzhen (China). It includes both abnormal
(containing marks of tuberculosis) and the standard CXR images. Unfortu-
nately, the dataset is not well-balanced in terms of absence or availability of

( A ) tuber- ( B ) pneu-
culosis monia
case case
F IGURE 2.8: Sample Chest X-Ray images containing marks of

tuberculosis (A) and pneumonia (B).
disease, gender, or age as presented on the chart 2.9. Here, we extracted only
153 samples of healthy patients (153 from both datasets) and 306 of those la-
beled with marks of tuberculosis. Selecting information about one class from
different resources ensures that the model does not learn features typical for
the method of taking images, e.g., lens.
Pneumonia is an inflammatory condition of the lung affecting the little
air sacs known as alveoli. Standard symptoms comprise of a blend of a dry

hack, inconvenience breathing, chest agony, and fever. The Labeled Optical
Tomography and Chest X-Ray Images for Classification dataset [48] include
selected images of patients from the Medical Center in Guangzhou. It con-
sists of data with two classes - normal and those containing marks of pneu-
monia. All data comes from the patient’s routine clinical care. The volume of
the complete dataset includes thousands of validated OCT and X-Ray images
yet for our analyze we wanted to keep the dataset tiny and evenly distributed
thus only 153 images were taken (another 153 images come from the tuber-
culosis dataset) from the resources labeled as healthy and 306 as pneumonia
- both chosen randomly. The exact dataset class distribution is presented in
F IGURE 2.9: "Combined distribution of genders and ages

among images without (0) and with (1) disease marks." [4]
Table 2.2
Healthy Pneumonia Tuberculosis

Number of samples 306 (153 per dataset) 306 306
TABLE 2.2: Dataset class distribution.
External segmentation of left and right lung images (exclusion of redundant
information; bones, internal organs, etc. - Fig. 2.10) was proven to be effec-
tive in gaining better prediction accuracy [4].
To extract lungs information and exclude outside regions, we used the man-
ually prepared masks included in the extension of the SH dataset, namely,

the segmented SH dataset. Due to nonidentical borders and lung shapes, the
segmentation data has high variability although its distribution is much sim-
ilar to the regular one, comparing to image area distribution as presented in

( A ) X-ray (B) ex-

image tracted
lungs
F IGURE 2.10: Example of an original image in SH dataset (left)

and the segmented result (right).
FIgure 2.11.
F IGURE 2.11: Distributions of image (Fig. 2.10, left) and mask

(Fig. 2.10, central) areas. [4]
2.6 Image Data Augmentation
Model-based methods greatly improve their predictions when increasing the
number of training samples. When a limited amount of data is available,

some transformations have to be applied to the existing dataset to syntheti-
cally increase the volume of the training set. Researchers in [44] incorporates
three techniques to augment the training dataset size. The first approach was
to randomly crop of a 224x224 pixel fixed-size window from a 256x256 pixel

image. The next technique was flipping the image horizontally, which al-
lowed capturing information about reflection invariance. Finally, the third

method added randomly generated lightning to capture color and lightning
variation. Image 2.12 shows sample operations on the image data.
(A) Original im- ( B ) Flipped image

age
(C) Darkened ( D ) Cropped im-

image (brightness age
change)
F IGURE 2.12: Selected image transformation methods. Image

A) shows the original image, B) flipped, C) darkened (bright-
ness change) and finally D) zoomed(cropped and upsized)
2.7 Evaluation
Evaluation of a machine learning model plays a crucial part in all projects.

One metric might not be satisfying when dealing with the effectiveness of
classifiers, and thus in this thesis, we will use the following metrics; accuracy,
F1-score, precision, sensitivity, specificity, a graphical performance - ROC (re-
ceiver operating characteristic curve) and AUC (Area under the ROC curve).
In this section, we define metrics for binary problems. To extend it for a three-
class problem we calculate them using the approach "one versus the others"
per every label [49].
Accuracy is the number of correctly classify objects to the total predictions:
Number o f Correct predictions

Accuracy = (2.7)
Total number o f predictions made
Now, let us consider a situation given in Table 2.3. We see that the data is
completely imbalanced, therefore predicting just the class neutral provides
us with the accuracy of 86% which is considered high. However, we are not
supposed to assume that the model is valid when the decision is based on
the dominating class. Therefore, we have to quantify our algorithms with

the following properties: recall and precision.
Class Number of instances

Healthy 600
Tuberculosis 50
Pneumonia 50
TABLE 2.3: Hypotetical data distribution
Recall (or sensitivity) R measures the number of classified samples with
respect to all the relevant ones. It calculates the ratio of true positives Tp to
the sum of false negatives Fn and true positives Tp .
Tp
R= (2.8)
Tp + Fn
Precision P explains how selected items are relevant. It is defined as the

ratio of true positives Tp to the sum of false positives and true positives:
Tp
P= (2.9)
Tp + Fp
That is to say; recall tells us how many samples we missed (in a positive
class). Precision tells us what proportion of positive samples were correctly
classified.
If one sample was diagnosed as tuberculosis (out of 50) and the rest as other,
then the precision P equals 100%. However, recall would be as low as 2%.
There is likewise specificity which is characterized as the proportion of true

negatives (Tn ) to the total of true negatives and false positives
Number o f True Negatives

Speci f icity =
Number o f True Negatives + Number o f False Positives
(2.10)
If a model reaches 100% of specificity, then it missed no True Negatives,

there were no False Positives - negative samples labeled as positive ones.
However, there is still a high risk of having False Negatives.

Using the example in Table 2.3, the specificity for the class Neutral is 0 know-
ing that there were no true negatives (Tn )

The F1-score is defined by the following formula:
Precission ⇤ Recall
F1 = 2 ⇤ (2.11)
Precission + Recall
This metric is a combination of recall and precision, keeping a balance be-
tween them.
Another way to show a classification model performance is a graph called
ROC (receiver operating characteristic curve) which creates a plot of the True
Positive Rate against False Positive Rate.
True Positive Rate (TPR) is just another name for Recall (R). False Positive
Rate (FPR) is defined by the following formula:
Fp
FPR = (2.12)
Fp + Tn
A ROC graph is a plot of True Positive Rate vs. False Positive Rate using vari-
ous classification thresholds. Increasing the threshold assigns fewer samples

as positive and thus decrease True Positives and False Positives. The reverse
situation takes place when lowering the classification threshold.

A typical ROC curve can be seen on image 2.13
In order to compute the values in a ROC curve, we use a sorting-based al-

gorithm called AUC. AUC means “Area under the ROC Curve,” and it mea-
sures the area underneath the two-dimensional curve from (0, 0) to (1, 1).
A typical ROC curve can be seen on image 2.14
Area Under the ROC Curve provides us with an aggregated measure of

F IGURE 2.13: Tp vs. FP using different thresholds of

classification.[5]
performance considering any classification threshold. It might be interpreted
as the probability that a given model assigns a random sample to a positive
class.
The area defined by AUC varies from 0 to 1. A model who predicts a cor-
rect class in 100% of cases has an AUC of 1.0 and another whose predictions
are 100% wrong has an AUC of 0.0.
This metric is desirable thanks to two reasons:
classification-threshold-invariance
scale-invariance
The first one measures how well model predictions are irrespective of
the chosen threshold. The second one estimates the rank of the predictions,
rather than their absolute values.
F IGURE 2.14: AUC (Area under the ROC).[5]
2.8 Convolutional Neural Networks
After publishing AlexNet ([44]) in 2012, convolutional neural networks (CNN)
renewed the interest of the research community. CNN and subsequent deep
models such as VGG ([43]) proved their usefulness, especially in computer

vision-related tasks. The contribution of published work demonstrates that
those models are suited much better at capturing different features than tra-
ditional algorithms which heavily rely on different feature engineering meth-

ods (e.g., gradient change).
In a classical formulation of a convolutional neural network used in classi-
fication, CNN consists of multiple convolution layers followed by pooling

operators.
The convolution layers are made of kernels, small tensors compared to win-
dows which process input and output information. Those operators can suc-
cessfully capture the spatial and temporal dependencies in an image and

thus learn different local features like straight lines (horizontal or vertical)
and curves while upper layers (hidden) can perform detection of more so-
phisticated information like rectangles or circles based on the received input,

therefore, understand it better. As processed data flows higher to deeper lay-
ers, a network learns more “abstractive” combinations.
F IGURE 2.15: Convolutional Neural Network for image

processing.[6]
Let us consider a hypothetical image presented in Figure 2.16 where we
can distinguish three color layers: red, green, and blue respectively. The role
of a convolutional neural network is to reduce the image space into a form
much more comfortable to process without loss of any information crucial

for obtaining a valid prediction. This example shows an image with small
dimensionality (4x4), yet this aspect gains its importance along with the in-
creasing size of an input, e.g., 32 Megapixel (6464 x 4864).

The example in Figure 2.17 shows the process of convolution operation,
which extracts valuable input features and processes this information to the
next level, whereas reducing the dimensionality.
Convolution can be viewed as a sequence of operations where a single

operation centered on a group of surrounding values results with a single
output o (provided we have only one kernel).

F IGURE 2.16: A hypothetical 4x4 image.
(A) (B) (C)
F IGURE 2.17: Convolution operation.
o = f Wx,j + b (2.13)
where W 2 R p,q is a weight matrix (kernel), p is the output size of the convo-
lution, q is the window size, f represents the non-linearity, and b is bias. Both
parameters b and W are shared across all inputs.
Similarly to the convolution layer, the Pooling operator is also responsible for
spacial size reduction and thus decreasing computational resources used in
processing data, albeit the pooling layer contains no parameters (there is no
channel). Rather, pooling operators are deterministic, normally ascertaining

either the most extreme or the average estimation of the components in the
pooling window. These activities are max pooling and average pooling. Fig-
ure 2.18 presents the extraction of dominant, rotational invariant information
- max pooling.
(A) (B) (C)
F IGURE 2.18: Max pooling operation.
The ConvNet effectively learns the relations between surrounding pixels through-
out an image. Thanks to the convolution, the input is mapped into a con-
strained, abstracted representation which results in the output describing

dominant features
2.9 Summary
This chapter briefly reviewed the work related to the problem: Extreme Learn-
ing Machines, Semi-Supervised Learning models, and Automatic CXR screen-

ing system. We also introduced previous deep learning methods used in
Chest X-Ray analysis, pulmonary disease datasets, image data augmentation

techniques, and different types of results measurements used in classifica-
tion. We briefed deep convolutional neural networks and explained the op-
erations they conduct in image data analysis.
30
Chapter 3
Transfer Learning in Lung Diseases

Classification
3.1 Transfer learning
Transfer learning is a method of optimization of the training process by using
tools pre-trained in a different task. A pre-trained model is reused as a base
for a different task.
“Transfer learning is the improvement of learning in a new task through
the transfer of knowledge from a related task that has already been learned.”
- Chapter 11: Transfer Learning, Handbook of Research on Machine Learn-
ing Applications, 2009.
This is a very popular approach in computer vision related tasks using

deep neural networks when data resources are limited. Therefore, to create a
starting point for a new task, we incorporate the pre-trained models skilled in
solving similar problems. This method is crucial in medical image processing
due to the shortage of sample volume.
In deep neural networks, feature extraction is conducted but passing raw

data through models specialized in other tasks. Here, we can refer to deep
Chapter 3. Transfer Learning in Lung Diseases Classification 31
learning models such as ResNet where the last layer information serves as
input features to a new classifier.
3.1.1 Pre-trained models approach
Transfer learning in deep learning problems can be performed using a com-

mon approach called pre-trained models to approach.
Here we can distinguish the three following approaches:
Reuse Model
True Model
Select Source Model
The first option, Reuse Model, states that a pre-trained model can produce
a starting point for another model used in a different task. This involves the
incorporation of the whole model or its parts.
In the second approach, an adopted model may or may not need to be
refined on the input-output data for the new task.
The third option considers selecting one of the available models. It is very
often that research institutions publish their algorithms trained on challeng-

ing datasets which may fully or partially cover the problem stated by a new
task.
3.1.2 ImageNet
ImageNet [50] is a project that helps computer vision researches in classi-
fication and detection tasks by providing them with a large image dataset.
This database contains roughly 14 million different images from over 20.000
classes. ImageNet also provides bounding boxes with annotations for over 1
million images, which are used in object localization problems.
In this work, we will focus on three models VGG, ResNet, and Inception
pre-trained on the ImageNet dataset.
3.2 VGG16
VGG model is a deep convolutional neural network proposed by researchers

A. Zisserman and K. Simonyan from the University of Oxford [43] containing
over 138 million parameters. This model was able to achieve 7.4% error rate
on the ImageNet dataset (see section 3.1.2). It improved the AlexNet [44]
network by changing the kernel size and instead of 11x11 and 5x5 filters in
the first two layers, it implemented multiple smaller ones 3x3 filters one after
another.
3.2.1 VGG16 Architecture
The VGG convolutional network is a model with 16 layers trained on fixed-
size images. The input is processed through a set of convolution layers which
use small-size kernels with a receptive field 3x3. This is the smallest size al-
lowing us to capture the notion of up, down, right, left, and center. The
architecture also incorporates 1x1 kernels which may be interpreted as linear
input transformation (followed by nonlinearity (see section 2.8). The stride
of convolutions (number of pixels that are shifted in every convolution - step

size) is fixed and set to 1 pixel; therefore the spatial resolution remains the
same after processing an input through a layer, e.g., the padding is fixed to
1 for 3x3 kernels. Spatial downsizing is performed by five consecutive pool-
ing (max-pooling) layers, which are followed by some convolution layers.

However, not all of them are followed by max-pooling. The max-pooling
operation is carried over a fixed 2x2 pixel window, with a stride of 2 pixels.
This pile of convolutional layers ends with three Fully-Connected (FC) layers
where the first two consist of 4096 channels each and the third one 1000 as it
performs the 1000-way classification using softmax. All hidden layers have
the same non-linearity ReLU(rectification) [44].
Figure 3.1 visualises the architecture of the VGG model with 16 layers.
F IGURE 3.1: VGG network with 16 layers.

3.3 ResNet-50
The ResNet convolutional neural network is a 50 layers deep model trained

on more than a million fix-sized images from the ImageNet dataset (see 3.1.2).
The network classifies an input image into one of 1000 object classes like car,
airplane, horse or mouse. The network has learned a plentiful amount of
features thanks to the training images diversity and can achieve a 6.71% top-
5 error rate on the ImageNet dataset (see section 3.1.2).
3.3.1 Residual Bloc
When expanding the number of layers in a convolutional neural system, a
normal phenomenon is to see a decreasing error. Unfortunately, the oppo-
site effect appears where accuracy saturates and eventually degrades. This,
however, is not caused by overfitting yet vanishing gradient [45].
Because of the effect related to vanishing, gradient, researchers were not
able to build deeper networks as they did not perform better than their shal-
lower counterparts. The main idea of the ResNet model is introducing an
“identity shortcut connection”(also residual bloc or skip connection) which

skips one or more layers. An example of such operation is presented in Fig-
ure 3.2
Let us consider a deep neural network block whose accurate output distri-
bution is denoted as H ( x ) transformation of input x. The following formula
defines the difference or the residue between those arguments:
R ( x ) = Output Input = H ( x ) x (3.1)

F IGURE 3.2: Residual bloc.
After rearrangement, we obtain:
H (x) = R (x) + x (3.2)
The residual block tries to learn the correct output H ( x ) and since there
is an identity connection there through skipping x to the output, the block
learns the residual R ( x ). Traditional networks try to predict output distri-

bution within their layers, whereas a residual deep convolutional neural net-
work learns the residual. Therefore, those blocks are called Residual.
3.3.2 ResNet50 Architecture
The ResNet-50 convolutional neural network is built of 5 stages, each hav-
ing convolutions and identity blocks. Every convolution block consists of 3

convolutional layers. Figure 3.4 shows an architecture of a slightly different
network (ResNet-34). However, the idea behind its sibling model remains
the same. The only difference is in residual blocks; unlike those in ResNet-34
(Figure 3.3 A) ResNet-50 replaces every two layers in a residual block with a
three-layer bottleneck block and 1x1 convolutions, which reduce and even-
tually restore the channel depth. This allows reducing a computational load
when a 3x3 convolution is calculated (Figure 3.3 B).
The model input is first processed through a layer with 64 filters each 7x7
and stride 2 and downsized by a max-pooling operation, which is carried
over a fixed 2x2 pixel window, with a stride of 2 pixels. The second stage
consists of three identical blocks, each containing a double convolution with

64 3x3 pixels filters and a skip connection block. The third pile of convolu-
tions starts with a dotted line (Figure 3.4) as there is a change in the dimen-
sionality of an input. This effect is achieved through the change of stride in
the first convolution bloc from 1 to 2 pixels. The fourth and fifth groups of
convolutions and skip connections follow the pattern presented in the third
stem od input processing, yet they change the number of filters (kernels) to
256 and 512, respectively[45]. This model has over 25 million parameters.
(A) (B)
Resnet-34 Resnet-50
residual residual
bloc. bloc.
F IGURE 3.3: Residual blocs
Visualisation of the network architecture can be found in Figure 3.4

F IGURE 3.4: Resnet-34 architecture

3.4 Inception
The researchers from Google introduced the first Inception (InceptionV1) [7]
neural network in 2014 during the ImageNet competition (See subsection
3.1.2). The model consisted of blocs called "inception cell" that was able to
conduct convolutions using different scale filters and afterward aggregate
the results as one. Thanks to 1x1 convolution which reduces the input chan-
nel depth the model saves computations. Using a set of 1x1, 3x3, and finally,
5x5 size of filters, an inception unit cell learns extracting features of different
scale from the input image. Although inception cells use max-pooling oper-
ator, the dimension of a processed data is preserved due to "same" padding,
and so the output is properly concatenated. A sample inception unit intro-

duced with InceptionV1 is presented in Figure 3.5
F IGURE 3.5: Inception cell introduced with the fist Inception

model[7]
A follow-up paper was released not long after introducing a more efficient
solution to the first version of the inception cell. Large filters sized 5x5, and
7x7 are useful in extensive spatial features extraction, yet their disadvantage
lies in the number of parameters and therefore computational disproportion.
3.4.1 Factorizing Convolutions
The researchers from Google found a way to save computations and reduce
number of parameters without dicreasing model’s efficiency. In the pro-
posed architecture [8] all 5x5 (Figure 3.6 (A)) convolutions were factorized
to two consecutive 3x3 (Figure 3.6 (B)) operations, improving the computa-
tional speed. New inception unit is presented in Figure 3.7
( A ) Dou- (B) 5x5

ble 3x3 convolu-
concolu- tion.
tion.
F IGURE 3.6: 5x5 convolution replaced by two 3x3 convolutions.
By processing input through a layer with a 5x5 pixel filter, the number of
parameters equals 25 (5x5 = 25). Whereas introducing two consecutive 3x3

pixel filters, the number of parameters decreases by 28% (2x3x3 = 18).
3.4.2 Factorization Into Asymmetric Convolutions
The researchers went even further with a decreasing number of filter param-
eters showing another double asymmetric convolutions 3x1 and 1x3 (see Fig-
ure 3.8) which deconstruct each 3x3 kernels.

F IGURE 3.7: Inception cell introduced in [8]
F IGURE 3.8: Asymmetric convolution [8]
Using deconstructed 3x3 filters the number of parameters decreases by roughly
33% since instead of 9 (3x3 = 9) we need only 6 (2x3x1 = 6). The application
of asymmetric convolution can be seen in Figure 3.9. To achieve the results
of a bloc presented in Figure 3.14, n needs to be set to 3.
The filter banks were furthermore expanded, making them wider, not
F IGURE 3.9: Inception cell using aymmetric convolutions[8]
deeper, which removed the representational bottleneck. Increasing the depth
of a module decreases the dimensionality and introduces information loss.
This is illustrated in Image 3.10
The suggested 3 different kinds of inception modules with factorization,

drastically reduces the number of parameters in the whole network. There-
fore, models incorporating such techniques are less prone to overfit and con-
sequently can get deeper.
3.4.3 Auxiliary Classifiers
First auxiliary classifiers (see image 3.11) were proposed along with the In-
ceptionV1 model [7]. Although the new models use the intuition behind
them, they are slightly modified. Instead of using 2 auxiliary classifiers [7],
only one is used on top of the 17x17 pixels layer (see Figure 3.14). The reason
F IGURE 3.10: Wider inception module[8]
for introducing this difference lies in their purpose. The first version of the
Inception deep neural network used auxiliary classifiers in order to make the
model deeper. Here, for instance, one classifier serves as a regularizer.
F IGURE 3.11: "Auxiliary classifier on top of the last 17x17 layer.

Batch normalization of the layers in the side head results in a
0.4% absolute gain in top-1 accuracy. The lower axis shows the
number of iterations performed, each with batch size 32." [8]
3.4.4 Effective Grid Size Reduction
The standard convention presented in AlexNet, VGG or ResNet uses a pool-

ing operator (see Section 2.8) in order to downsize the feature map (See image
3.12). However, the drawback of this approach is either having a computa-

tionally expensive convolution operation followed by pooling or a greedy by
max-pooling procedure proceeding with a convolution layer. Therefore, an

alternative solution proposed effective grid size reduction.
F IGURE 3.12: "Two alternative ways of reducing the grid size.

The solution on the left violates principle 1 of not introducing a
representational bottleneck from Section 2. The version on the
right is 3 times more expensive computationally." [8]
The output from the effective grid size-reduction bloc is a concatenation of
two sets of feature maps, together with having 640 channels. The first set
having 320 feature maps is an output from a convolution bloc with stride
equal 2. The second set constituting another 320 channels is obtained by max
pooling.
The effective grid size reduction is an efficient operation, although less
expensive.
The Figure 3.13 an inception module redicing grid-size.
3.4.5 Architecture
The InceptionV3 model [8] contains over 23 million parameters. The ar-
chitecture can be divided into 5 modules, as presented in Figure 3.14. The
F IGURE 3.13: "Inception module that reduces the grid-size

while expands the filter banks. It is both cheap and avoids the
representational bottleneck as is suggested by principle 1. The
diagram on the right represents the same solution but from the
perspective of grid sizes rather than the operations." [8].
first processing block consists of 3 inception modules visualized in the im-
age 3.6. Then, information is passed through the effective grid size reduction
(see 3.4.4) and processed through four consecutive inception cells with asym-
metric convolutions (see image 3.8). Moving forward, information flows to

the 17x17 pixels convolution layer connected to an auxiliary classifier (see
3.4.3) and another effective grid size-reduction block. Finally, data progresses
through a series of two blocs with wider filter banks (see image 3.10) and con-
sequently gets to a fully-connected layer ended with a Softmax classifier.
Visualization of the network architecture can be found in Figure 3.14
F IGURE 3.14: InceptionV3 architecture. Batch normalization

and ReLU non-linearity are used after every convolution layer.)
3.5 Experiments
3.5.1 Dataset
The first part of the experiments compares three modified versions of neural
networks introduced earlier; VGG16, ResNet-50 and InceptionV3 described
in 3.2, 3.3 and 3.4, respectively. We train those models on the database con-
taining X-Ray lungs images introduced in Section 2.5. All images 2.2 (918
samples, 306 per class) were resized to the same shape before training, 256x256
pixels. The dataset constituting of X-Ray images and one-hot encoded labels
was partitioned into three different categories for training, validation, and
testing, respectively. The training dataset consisted of 80% of randomly se-

lected images. The validation set used 10% of all data, and the remaining
10% served for testing (the same approach was used in [4]), and its class
distribution was kept in an even proportion, e.g., a third of samples were la-
beled as ’healthy’, a third contained marks of tuberculosis and the remaining
part came from patients suffering from pneumonia. During the training pro-
cess, the input data is augmented [51] [52][53] by randomly selecting one of
the following operations: rotation, brightness change, and random cropping

(See section 2.6). The validation set used to control the accuracy and overfit-
ting to training data is created independently in each training using Monte
Carlo Cross Validation [54].
3.5.2 Models
The analysis in this chapter compares three models using different transfer
learning described in 3.2, 3.3 and 3.4 for lung disease classification. The mod-
els were expanded with the same neural networks based classifier consisting
of a global average pooling layer, three fully connected layers having 1024,
512 and 256 neurons and a softmax classifier. The number of trainable param-
eters in the deep neural networks is 1,182,211 for VGG16 and 2,755,075 for
both ResNet50 and InceptionV3. Before passing input images through deep
neural networks which serve as feature extractors, the batches were adjusted
to the same formats (batch size, input scale, etc.) the mentioned models were
trained with [43][45][8]. Furthermore, our neural network-based classifiers

learn bias and weights parameters by backpropagating the error to minimize
the categorical cross-entropy using Adam [55] optimizer at training time.
The models were later on validated using different sets of hyper-parameters

(only appended layers as pre-trained networks remain frozen) observed dur-
ing the training process. The output from the softmax classifier is a vector of
probabilities with which an input image belongs to one of the classes. The
final class is the one corresponding to the highest value, and its position is
then mapped back to a corresponding class.
The code for the transfer-learning models is publicly available through a
python API, Keras. Our algorithms were trained on servers equipped with
GPU provided by Helios Calcul Québec, which consists of fifteen computing
nodes each having eight Nvidia K20 GPUs and additionally six computing
nodes with eight nVidia K80 boards each. Every K80 board includes two
GPU’s and so the total of 216 GPU’s in the cluster.
3.6 Results and analysis
Training three models repeatedly ten times for 150 epochs took roughly one
day. This relatively short period is caused by setting parameters of pre-
trained networks to non-trainable, and thus the gradient caused by misclas-
sification flows only through the appended layers. The initial image prepro-
cessing was also advantageous for the duration since the real size images of
hundreds of thousands or million pixels were initially downsized to a fixed

format. The biggest problem with training was related to the maximum plat-
form usage time, which is up to twelve hours.

The examination of error calculated on a validation set allowed us to find
a relatively good epoch after which models were overfitting training data.
The training process was then stopped, and the final results were measured
as an average of all results obtained at that step. The last step was to show
the performance of selected models on the unseen data (test set).
3.6.1 VGG results
The following results were generated for ten independent training runs to
observe a similar training pattern. Each of the ten training and validation
curves (see Figures 3.15 and 3.16) were plotted on the same charts based on
the tape (training or validation). To maintain a high level of readability, the
results were separated. The wider, dotted curve is an averaged result of all
obtained at the particular epoch. The red dot in Figures in 3.15 represent
the lowest loss value on training and validation data sets, whereas in 3.16 it
corresponds to the maximum accuracy obtained.

Figure 3.15 shows that the model slowly starts overfitting on the training
dataset around the 90th epoch yet then the validation error falls again and
eventually after 150 epoch achieves the best average results. Similar behavior
is experienced when examining Figure 3.16. Here, the average validation

accuracy slows down but then slightly increases.
Eventually, the models were evaluated on the test set and scored an av-
erage accuracy of 63.85% In order to visualize results, we selected a network

which obtained the best accuracy score. We see on the confusion matrix
in Figure 3.17 A) that the model had the biggest problems with classifying
’tuberculosis images’ to the corresponding class and tend to mistake it as

( A ) Training error change
( B ) Validation error change
F IGURE 3.15: VGG16 based model training and validation loss

change.
( A ) Training accuracy change
( B ) Validation accuracy change
F IGURE 3.16: VGG16 based model training and validation ac-

curacy change.
healthy. Fortunately, the majority of samples containing marks of tubercu-

losis were correctly classified. The image B) in Figure 3.17 shows that AUC
score for classes healthy, tuberculosis and pneumonia were equal 0.68, 0.82
and 0.80, respectively. Additionally, we present sample classification results
in Figure 3.18. Although both images A) and C) were correctly classified,

their corresponding class activation maps (images B) and D)) show that their
determinative regions were not related to the problem. Instead of investi-

gating lung regions, our trained model focused on waist curves, armpits, or
collarbones.
( A ) Confusion matrix
( B ) Per class ROC curves
F IGURE 3.17: Image A shows the confusion matrix of the ob-

tained results with the model which reached the highest ac-
curacy during the training. Image B show its per class ROC
curves.
( A ) Correctly clas- (B) Class activa-

sified image as ‘tu- tion map
berculosis‘
(C) Correctly ( D ) Class activa-

classified image as tion map
‘pneumonia‘
F IGURE 3.18: Two pairs of correctly labelled images containing

marks of tuberculosis and pneumonia with their class activa-
tion maps.
3.6.2 ResNet results
The following results were generated for ten independent training runs in
order to observe a similar pattern in model behavior during training. All

the results achieved during training and validation processes were plotted in
Figures 3.19 and 3.20. By splitting the results by type (separately training and
validation), we maintained a high level of visibility, allowing us to simplify
the analysis. The wider, dotted curve is an average of all results obtained
at the corresponding epoch. The red dot in Figures 3.19 and 3.20 stands for
minimum loss and maximum accuracy value, respectively. The Figure 3.19
shows that the model decreases its loss throughout the whole training yet
seems to reach the validation accuracy plateau around 120th epoch (see Fig-
ure 3.20 B)).
Finally, after 150 epochs, all the ResNet-50 based models were evaluated
on the test set and scored an average accuracy of 72.22%, which is almost
a ten points improvement to 3.6.1. In order to visualize the results, we se-
lected a network which achieved the best accuracy score. The confusion ma-
trix in Figure 3.21 A) shows that, similarly to 3.6.1, the model had problems
with correctly labelling ‘tuberculosis images‘ although the majority of images
were correctly classified. Image B) shows that the AUC score for healthy, tu-
berculosis and pneumonia were equal 0.84, 0.76, and 0.84, respectively. This
is an improvement in comparison to results in the previous subsection (3.6.1).
Additionally, we present sample classification results in Figure 3.22. Simi-
larly to 3.6.1, both images A) and C) were correctly classified yet the overlap-
ping class activation maps (images B) and D)) show that the determinative
regions were not related to lungs. Areas like collarbones or heart decided of
the final label.
F IGURE 3.19: ResNet-50 based model training and validation

loss change.

accuracy change.

curves.

berculosis‘

‘pneumonia‘

tion maps.
3.6.3 Inception results
order to observe a similar training pattern. Each of the ten training and val-
idation curves (see Figures 3.23 and 3.24) were plotted on the same charts
based on the tape (training or validation). The same as before, we separate

the results to maintain a high level of readability. The wider, dotted curve
is an averaged result of all ten runs at the particular epoch. The red dot in
Figures 3.23 3.24 stand for the lowest loss value on training and validation
data sets the maximum accuracy obtained, respectively.

The image 3.23 shows that the model starts overfitting on the training
dataset around the 20th epoch which can be witnessed by examining both
validation error and accuracy change. The validation error curve slowly in-
creases its value, whereas the accuracy level remains similar (see Figure 3.24).
Finally, we took all InceptionV3 based models at the 20th epoch and evalu-
ated them on the test set and scored an average accuracy of 80.55%, which
is a significant improvement comparing to 3.6.1 and 3.6.2. To visualize the

final results, similarly to the previous chapter, we selected a model which
achieved the best accuracy score after the 20th epoch. The confusion matrix
in Figure 3.25 A) shows that the new model improved the number of true
positives (TP) in all classes comparing to previous algorithms, although the
network still faced minor problems with classifying all ‘tuberculosis images‘
to their corresponding label. Image B) shows that AUC score for healthy,
tuberculosis and pneumonia were equal 0.89, 0.87, and 0.92, respectively.
This is an improvement in comparison to results in the previous subsections

(3.6.1).
Additionally, we present sample classification results in Figure 3.26. Sim-

ilarly to previous subsections (3.6.1 and 3.6.2), both images A) and C) were
correctly classified yet the overlapping class activation maps (images B) and
D)) show that the determinative regions were not necessarily related to lungs.
Areas like armpits and internal organs decided on the final label.
F IGURE 3.23: InceptionV3 based model training and validation

loss change.

accuracy change.

curves.

berculosis‘

‘pneumonia‘

tion maps.
3.6.4 Results comparison
After comparing the results obtained in sections 3.6.1, 3.6.2 and 3.6.3 we can
observe that transfer-learning models perform decently in lung diseases clas-

sification tasks even when the data resources are limited. Table 3.1 shows the
comparison of all trained algorithms. The InceptionV3 based model scored

the best, reaching better accuracy than VGG16 algorithms by over 16 points.
What is more, we see major improvement in AUC, F1 score, precission and

sensitivity. The most important results are related to the "infected" classes
and here we see that using InceptionV3 model a random instance contain-
ing marks of either tuberculosis or pneumonia has a high probability to be
assigned to one of those classes. The interpretability of our methods is not

certain because the generated class activation maps show that our models
focused on regions outside the lungs area - hips or collarbones. However,

we can clearly state that using transfer-learning based algorithms on small
datasets allows achieving competitive classification scores on the unseen data.
There was no work done with a similar dataset (Chest X-Ray multiclassifi-
cation problem with small dataset) therefore we did not compare ourselves
with available solutions.
VGG16 ResNet-50 InceptionV3

Accuracy 0.64 0.72 0.81
AUC (healthy) 0.68 0.84 0.89
AUC (pneumonia) 0.80 0.84 0.92
AUC (tuberculosis) 0.82 0.76 0.87
F1 score (healthy) 0.55 0.66 0.76
F1 score (pneumonia) 0.64 0.82 0.90
F1 score (tuberculosis) 0.72 0.66 0.75
precision (healthy) 0.49 0.77 0.80
precision (pneumonia) 0.87 0.76 0.89
precision (tuberculosis) 0.71 0.67 0.74
sensitivity (healthy) 0.62 0.60 0.73
sensitivity (pneumonia) 0.53 0.88 0.92
sensitivity (tuberculosis) 0.77 0.68 0.77
TABLE 3.1: Comparison of transfer-learning based algorithms

in terms of accuracy, AUC, F1 score, precision and sensitivity
for healthy, pneumonia and tuberculosis classes.
3.7 Summary
This chapter introduces three models which achieved the highest scores in
the ImageNet competition; VGG16, ResNet-50, and InceptionV3. We also
present our initial work in lung diseases classification using those pre-trained
deep neural networks as feature extractors for a simple 3-layers deep neural
network. The significant and promising results on small datasets show that
there is no need to build sophisticated, multiple-layers networks in order to
achieve high scores on the test dataset.

66
Chapter 4
Transfer Learning Models Accuracy

Improvement in Lung Diseases
Classification Using Segmentated
X-Ray Images
4.1 U-Net - Image Segmentation using Deep Neu-
ral Networks
Many vision-related tasks, especially those from the field of medical image
processing, expect to have a class assigned per pixel, i.e., every pixel is as-
sociated with a corresponding class. To conduct this process, we propose a
neural network architecture described in [9] and showed in Figure 4.1.

This model works well, as proven in [9] and [4], with very few training image
examples yielding precise segmentation. The thought behind this network is
to utilize progressive layers instead of a building system, where upsampling

layers are utilized as instead of pooling operators. Therefore, increasing the
output resolution.
High-resolution features are combined with the upsampled output to do lo-
calize, as presented in Figure 4.1. The deconvolution layers consist of a high

Chapter 4. Transfer Learning Models Accuracy Improvement in Lung
67
Diseases Classification Using Segmentated X-Ray Images
number of kernels, which better propagate information and result in higher

resolution output. Thanks to the described procedures, the deconvolution
path is approximately symmetric to the contracting one, and so the archi-

tecture resembles a u-shape. There are no fully connected layers, therefore,
making it possible to conduct the seamless segmentation of relatively large

images, extrapolating the missing context by mirroring processed input.
F IGURE 4.1: "U-net architecture (example for 32x32 pixels in

the lowest resolution). Each blue box corresponds to a multi-
channel feature map. The number of channels is denoted on
top of the box. The x-y-size is provided at the lower left edge
of the box. White boxes represent copied feature maps. The
arrows denote the different operations."[9]
4.1.1 Architecture
The network architecture showed in Figure 4.1 consists of an expansive path
(right) and a contracting one(left). The first part (contacting)resembles a typi-

cal convolutional neural network; the repeated 3x3 convolutions followed by
a non-linearity, here rectified linear unit (ReLU), and 2x2 poling with stride
2. Each downsampling operation doubles the number of resulting feature
maps.
All expansive path operations are made of upsampling of the feature chan-
nels followed by a 2x2 deconvolution (or "up-convolution") which reduces
68
the number of feature maps twice. The result is then concatenated with the
corresponding feature layer from the contracting path and convolved with
3x3 kernels, and each passed through a ReLU. The final layers apply a 1x1
convolution to map each feature vector to the desired class.
69
4.2 Lungs Segmentation
Following the approach presented by previous works summarized in Chap-

ter 2, we wanted to use deep convolutional neural networks to segment lungs
[56] before processing it through the classification models mentioned in 3.5.2.

Researchers in [56] indicate that U-Net architecture and its modifications out-
perform the majority of CNN based models and achieve excellent results by
easily capturing spacial information about the lungs. As an outcome, we
propose a pipeline that consists of two stages; first segmentation and then
classification.
4.2.1 Dataset
The phase of extracting valuable information (lungs) is conducted with a
model presented in 4.1. Our algorithms trained for 500 epochs on an ex-
tension of the SH dataset described in 2.5. The input to our u-shaped deep
neural network is a regular Chest X-Ray image, whereas the output is a man-
ually prepared binary mask of lung shape, matching the input. Figure 4.2 A)
presents an X-Ray image and B) its corresponding mask.
4.2.2 Training
As mentioned before, our model was trained for 500 epochs using a dataset
divided into 80%, 10%, and 10% parts, for training, validation and test parts
respectively on the same machines introduced in section 3.5.2 using the batch
size o 8 samples, augmentation techniques briefed in subsection 2.6, Adam

optimizer [55] and categorical cross-entropy as a loss function for pixel-wise
binary classification. The training results are visible in Figure 4.3. As we
can easily notice, the validation error is slowly falling throughout the whole
training, whereas there is no major change after the 100th epoch. The final
70
( A ) X-Ray image ( B ) Correspond-

ing lungs mask
F IGURE 4.2: An X-Ray image and its corresponding lungs

mask.
error on the validation set is right below 0.05 and slightly above 0.06 on the
test set.
4.2.3 Results
Our algorithm learns shape-related features typical for lungs and can gener-
alize well further over unseen data. Figure 4.4 shows the results of our U-Net
trained models.
It is clear that the network was able to learn chest shape features and ex-
clude regions containing internal organs such as the heart. The incredibly
promising results allowed us to process the whole dataset presented in sub-
section 3.5.1 and continue our analysis on the newly processed images.
71
F IGURE 4.3: U-Net training and validation losses change dur-

ing training.
72
( A ) X-Ray image (B) Segmented

lungs
( C ) X-Ray image (D) Segmented

lungs
( E ) X-Ray image (F) Segmented

lungs
F IGURE 4.4: Three pairs of CXR images with corresponding,

segmented (extracted) lungs.
73
4.3 Training Deep Learning Models On Segmented
Images
As disclosed in the prologue to this section, we propose a two-stage pipeline

consisting of Chest X-Ray images segmentation and lung disease classifica-
tion. The first phase (segmentation) is conducted based on experiments pre-

sented in the previous section 4.2. The second stage is to train the models
presented in the previous chapter 4 to investigate the potential improvement

in performance. Our classification models were trained using the same setup
as shown in 3.5.2
4.3.1 Dataset
Here, we conduct our experiments using the same data as in chapter 3. The
difference is in previous segmentation, which extracts valuable for the task
information - lungs. Figure 4.4 shows the training samples; the left and right
column correspond to input and output, respectively.
4.4 Results and analysis
As in the previous chapter, training all three models repeatedly ten times
for 150 epochs took about one day. This short time is a result of setting all
parameters of pre-trained models as non-trainable. Therefore the gradients
flow only through the concatenated layers. The segmented images were re-
sized, which also beneficially influences the execution duration. The biggest
problem related to training our models was the maximum platform usage
time, which is up to twelve hours.

The examination of training and validation error curves allowed us to
find a relatively good number of epochs after which models were overfitting
74
the training data. The training process was then stopped, and the final results
were measured as an average of all results obtained at that step.
4.4.1 VGG results
The following results were generated for ten independent training runs to
observe a similar training pattern. Each of the ten training and validation
curves (see Figures 4.5 and 4.6) were plotted on the same charts based on
the tape (training or validation). So as to keep up a high state of readability,
the results were isolated. The wider, dotted curve is an averaged result of
all obtained at the particular epoch. The red dot on Figures in 4.5 represent
the lowest loss value on training and validation data sets, whereas on 4.6 it
corresponds to the maximum accuracy obtained.

The image 4.5 shows that the model stops dropping the validation er-
ror around the 70th epoch and maintains it’s level throughout the remaining
roughly 80 epochs. A similar behavior is experienced when examining Fig-
ure 4.6. Here, the average validation accuracy slows down and only slightly
increases.
Eventually, the models were evaluated on the test set and scored an aver-
age accuracy of 69.99%, which is over 6 percentage of improvement compar-
ing to the results received in 3.6.1. To visualize results, we selected a network
which obtained the best accuracy score after 150 epochs. The confusion ma-
trix in Figure 4.7 A) shows that the model had the biggest problems with clas-
sifying ’pneumonia images’ to the corresponding class and tend mistaken it
as healthy. The other problem is related to labeling healthy images as healthy,

although in this case, it is not a major issue since the cost of assigning healthy
patients to a category of sick ones is more acceptable than otherwise. Addi-
tionally, the majority of samples containing marks of pneumonia were cor-

rectly classified. The image B) on Figure 4.7 shows that AUC score for classes
75
F IGURE 4.5: VGG16 based model training and validation loss

change.
76
F IGURE 4.6: VGG16 based model training and validation accu-

racy change.
77
healthy, pneumonia and tuberculosis were equal 0.75, 0.81 and 0.90, respec-
tively. This is also an improvement with the results received in 3.6.1.
Additionally, we present sample classification results on the Figure 4.8. Both
images A) and C) were correctly classified and corresponding class activation

maps (images B) and D)) show that their determinative regions were related
to the problem, unlike in 3.6.1. The network investigated lungs regions and
made the final decision based on extracted features.
78

curves.
79

berculosis‘

‘pneumonia‘

tion maps.
4.4.2 ResNet results
order to observe a similar pattern in model behavior during training. All

the results achieved during training and validation processes were plotted in
Figures 4.9 and 4.10. By splitting the results by type (separately training and
validation), we maintained a high level of visibility, allowing to simplify the
80
analysis. The wider, dotted curve is an average of all results obtained at cor-
responding epoch. The red dot on Figures 4.9 and 4.10 stands for minimum
loss and maximum accuracy value, respectively. The image 4.9 shows that
the model decreases its loss throughout the whole training.
Finally, after 150 epochs all the ResNet-50 based models were evaluated
on the test set and scored an average accuracy of 74.99%, which is almost a 5
percent improvement to 4.4.1 and over two to 3.6.2. In order to visualize the
results, we selected a network which achieved the best accuracy score. The
confusion matrix in Figure 4.11 A) shows that, similarly to 3.6.2, the model
had problems with correctly labelling ‘tuberculosis images‘ although the ma-
jority of images were correctly classified. Image B) shows that AUC score for
healthy, pneumonia and tuberculosis were equal 0.77, 0.91, and 0.82, respec-
tively. Comparing to the results in the previous subsection 3.6.1, we were
able to improve the results for healthy and pneumonia images classification.
When looking at AUC scores in 3.6.2, we scored better in labeling pneumo-
nia.
Additionally, we present sample classification results on the Figure 4.12. Sim-
ilarly to 4.4.1, both images A) and C) were correctly classified and the over-
lapping class activation maps (images B) and D)) show that the determinative
regions were related to lungs.
81

loss change.
82

accuracy change.
83

curves.
84

berculosis‘

‘pneumonia‘

tion maps.
4.4.3 Inception results
order to observe a similar training pattern. Each of the ten training and val-
idation curves (see Figures 4.13 and 4.14) were plotted on the same charts
based on the tape (training or validation). The same as before, we separate

the results to maintain a high level of readability. The wider, dotted curve
85
is an averaged result of all ten runs at the particular epoch. The red dot on
Figures 4.13 4.14 stand for the lowest loss value on training and validation
data sets the maximum accuracy obtained, respectively.

Figure 4.13 shows that the model starts overfitting on the training dataset
around the 40th epoch, although we do not witness any noticeable drop
in terms of accuracy. The validation error curve slowly increases its value,
whereas the accuracy level remains similar (see Figure 4.14). Finally, we
took all InceptionV3 based models at the 40th epoch and evaluated them on
the test set and scored an average accuracy of 82.22%, which is a small im-
provement comparing to results obtained using non-segmented X-Ray im-

ages in3.6.3. However, this experiment provided us with the best average
accuracy across all the models we trained using different techniques. In or-
der to visualize the final results, similarly to the previous chapter, we se-
lected a model which achieved the best accuracy score after the 40th epoch.
The confusion matrix in Figure 4.15 A) shows that the new model improved
the number of true positives (TP) in all classes comparing to previous algo-
rithms. Image B) shows that the AUC score for healthy, tuberculosis and
pneumonia were equal to 0.90, 0.93, and 0.99, respectively. This is a slight
drop for the healthy class comparing to results in 4.4.1 and 3.6.3 albeit we
greatly improved in comparison to our first results in 3.6.1.
Additionally, we present sample classification results in Figure 4.16. Sim-

ilarly to previous subsections (4.4.1 and 4.4.2), both images A) and C) were
correctly classified and the overlapping class activation maps (images B) and
D)) show that the determinative regions were related to langs.

86

loss change.
87

accuracy change.
88

curves.
89

berculosis‘

‘pneumonia‘

tion maps.
90
4.5 Comparison of results
After comparing the results obtained in subsections 4.4.1, 4.4.2 and 4.4.3 we
can observe that transfer-learning models perform well in lung diseases clas-
sification using segmented images tasks even when the data resources are
limited. Not only is their accuracy improved, yet also the class activation
maps support our conclusion. Table 4.1 shows comparison of results for all
trained algorithms, both using segmented and non-segmented Chest X-Ray
mages.
Non segmented CXRs Segmented CXRs

VGG16 ResNet-50 InceptionV3 VGG16 ResNet-50 InceptionV3
Accuracy 0.64 0.72 0.81 0.70 0.75 0.82
AUC (healthy) 0.68 0.84 0.89 0.75 0.77 0.90
AUC (pneumonia) 0.80 0.84 0.92 0.81 0.91 0.99
AUC (tuberculosis) 0.82 0.76 0.87 0.90 0.82 0.93
F1 score (healthy) 0.55 0.66 0.76 0.55 0.62 0.76
F1 score (pneumonia) 0.64 0.82 0.90 0.84 0.89 0.93
F1 score (tuberculosis) 0.72 0.66 0.75 0.79 0.73 0.78
precision (healthy) 0.49 0.77 0.80 0.67 0.66 0.75
precision (pneumonia) 0.87 0.76 0.89 0.92 0.93 0.90
precision (tuberculosis) 0.71 0.67 0.74 0.68 0.71 0.81
sensitivity (healthy) 0.62 0.60 0.73 0.48 0.62 0.77
sensitivity (pneumonia) 0.53 0.88 0.92 0.78 0.85 0.95
sensitivity (tuberculosis) 0.77 0.68 0.77 0.95 0.78 0.75
TABLE 4.1: Comparison of all results received on segmented

and non-segmented data.
The algorithm that scored the best in the majority results was InceptionV3
trained on the segmented images. What is more, it produced incredibly high
scores for the "diseased" classes showing that a random instance containing
marks of tuberculosis or pneumonia has over 90% probability to be classified
to the correct class. Although the scores of the healthy class are worse than
the diseased ones, its real cost is indeed lower as it is always worse to classify
a sick patient as healthy.

The InceptionV3 based model scored the best, reaching better accuracy than
91
VGG16 algorithms by over 12 percentage. Although the interpretability of

our methods is not guaranteed, we can clearly state that using transfer-learning
based algorithms on small datasets allows achieving competitive classifica-

tion scores on the unseen data.
(A) Activation (B) Activation

class map over a class map over a
non segmented segmented image
image
(C) Activation (D) Activation

class map over a class map over a
non segmented segmented image
image
F IGURE 4.17: Two pairs of correctly classified images with their

class activation maps. The left columns is non segmented im-
ages and the right is the segmented ones.
92
Furthermore, we compared the class activation maps and in order to inves-

tigate the reasoning behind decision making. The remaining features, here
lungs, force the network to explore it and thus make decisions based on ob-
served changes. That behavior was expected and additionally improved the
interpretability of our models as the marked regions might bring attention in

case of sick patients.
4.6 Comparison with other works
In this section, we would like to compare our models to the results achieved
over different datasets. In order to do so we trained our algorithms on Shen-
zhen and Montgomery datasets [11] ten times, generated the results for all
the models and averaged their scores: accuracy, precision, sensitivity, speci-
ficity, F1 score, and AUC 2.7
Table 4.2 shows the comparison of different deep learning models trained
on the Shenzhen dataset [11]. Although our approach does not guarantee the
best performance, it’s close to the highest one yet less complex. Researchers
in [57] use various pre-trained models in the pulmonary disease detection

task, and the ensemble of them presents the highest accuracy and sensitivity.
To compare, our InceptionV3 based model achieves accuracy smaller by only
one percent and equal AUC, which means that our method gives an equal
probability of assigning a positive case of tuberculosis to its corresponding

class over a negative sample. Although we could not outperform the best
solution, our approach is less complicated.
Furthermore, we compared the performance of our approach trained on

the Montgomery dataset [11] (see Table 4.3). Our InceptionV3 based model
tied with [12] in terms of accuracy yet showed a higher value of AUC. ResNet-
50 and VGG16 based models performed worse, however not drastically as

93
Accuracy Precision Sensitivity Specificity F1 score AUC

[58] 0.82 - - - - -
[12] 0.84 - - - - 0.90
VGG16 [57] 0.84 - 0.96 0.72 - 0.88
ReNet-50 [57] 0.86 - 0.84 0.88 - 0.90
ResNet-152 [57] 0.88 - 0.80 0.92 - 0.91
Ensemble [57] 0.90 - 0.88 0.92 - 0.94
VGG16 0.84 0.88 0.80 0.89 0.83 0.86
ResNet-50 0.85 0.97 0.73 0.98 0.83 0.92
InceptionV3 0.89 0.96 0.80 0.97 0.88 0.94
TABLE 4.2: Comparison of different deep learning based solu-

tions trained on the Shenzhen datases. Although our result is
not the best, it performs better than any single model (exclud-
ing Ensemble). Horizontal line means that those results were
not provided in literature.
Accuracy Precision Sensitivity Specificity F1 score AUC

[12] 0.790 - - - - 0.811
[59] 0.674 - - - - 0.884
[13] 0.783 - - - - 0.869
VGG16 0.727 0.842 0.581 0.872 0.669 0.931
ResNet-50 0.764 0.814 0.691 0.836 0.744 0.891
InceptionV3 0.790 0.822 0.745 0.836 0.779 0.884
TABLE 4.3: Comparison of different deep learning based so-

lutions trained on the Montgomery dataset [11]. Our average
performance is almost identical to [12].
they reached accuracies of 76% and 73% respectively, which is roughly 3 and
6 percent less than the highest score achieved.

94
Chapter 5
Conclusions and Future Work
5.1 Summary
In this thesis, we focused on exploring lung disease classification problem
using deep neural networks preceded by segmentation and not under the
supervision of the small size dataset (less than 103 examples). Moreover,
we examined class activation maps to explore the reasoning of our models
and investigate which regions are determinative. In Chapter 3 we first in-

troduced different deep learning architectures which competed in ImageNet
challenge [50]. Then, we use those networks as feature extractors to train
our shallow algorithms. The results are summarized in section 3.6.4, here
we only use accuracy to evaluate the performance since the test set is class
equally distributed. Chapter 4 introduces the U-Net deep neural network

and proposes a disease classification pipeline where Chest X-Ray images are
first segmented before processing them through models described in Chapter

3. We train the same algorithms using a preprocessed dataset and compare
the results with those obtained after learning features from non-segmented
images. Here we also show how our solutions outperform deeper models
trained on the same data. After comparing class activation maps in section
4.5, we conclude that segmentation not only improved the accuracy score yet
also the reasoning behind the classification. Preprocessed Chest X-Ray im-
ages with remaining lungs force networks to explore only those areas which
Chapter 5. Conclusions and Future Work 95
lead to improvements in the interpretability of our models as the marked

areas might bring attention in case of sick patients.
5.2 Contributions
The main contributions of this thesis are:
1. Implementation of transfer learning-based models such as VGG16, ResNet-
50, and InceptionV3 in lung disease classification tasks.
2. Implementation of the U-Net model for lung segmentation.
3. Evaluation of algorithms trained on segmented and non-segmented
Chest X-Ray images. Here, we compared both models’ accuracies and
their class activation maps.
4. Evaluation of the Shenzhen Hospital X-ray and Montgomery datasets
for tuberculosis prediction and comparison to previous work done with

deep neural networks.
5.3 Future Work
In Chapter 3, we examined only a small portion of available, pre-trained
networks in the classification task. Even though the results are promising,
other networks could be explored and examined in terms of class activation
maps. Furthermore, we only used one classifier (3 layers deep neural net-
work) due to computational and time limitations. Another direction would
be the application of the introduced solutions to much bigger datasets such as
ChestX-ray14 [60]. In Chapter 4, we propose a pipeline where classification

is proceeded by segmentation. This part opens another area of exploration.
As our models label Chest X-Ray images based on features extracted from
Chapter 5. Conclusions and Future Work 96
segmented data, yet it is beyond our expertise to decide whether the deter-
minative regions truly contain marks of disease.
This work presents only a small part of research on lung disease classi-
fication. Considering the promising results we achieved and the relatively
recent interest of deep neural network techniques in the medical field, there
is plenty of room for improvements in other biomedical applications. Fur-
thermore, we hope that one day computers will accelerate and help with a
radiological examination and save the lives of millions.
97
Bibliography
[1] A. Karargyris, J. Siegelman, D. Tzortzis, S. Jaeger, S. Candemir, Z. Xue,

K. Santosh, S. Vajda, S. Antani, L. Folio, and G. R Thoma, “Combination
of texture and shape features to detect pulmonary abnormalities in dig-

ital chest x-rays,” International Journal of Computer Assisted Radiology and
Surgery, vol. 11, pp. 563–576, June 2015.
[2] I. Livieris, A. Kanavos, V. Tampakas, and P. Pintelas, “An ensemble ssl
algorithm for efficient chest x-ray image classification,” Journal of Imag-
ing, vol. 4, p. 95, July 2018.
[3] S. Lopez-Garnier, P. Sheen, and M. Zimic, “Automatic diagnostics of tu-
berculosis using convolutional neural networks analysis of mods digital
images,” PLOS ONE, vol. 14, pp. 1–16, 2019.
[4] S. Stirenko, Y. Kochura, O. Alienin, and O. Rokovyi, “Chest x-ray anal-

ysis of tuberculosis by deep learning with segmentation and augmenta-
tion,” 2018 IEEE 38th International Conference on Electronics and Nanotech-
nology (ELNANO), pp. 422–428, 2009.
[5] Google, “Machine learning crash course with tensorflow apis.” https:
//developers.google.com/machine-learning/crash-course/.
[6] Y. LeCun and Y. Bengio, The Handbook of Brain Theory and Neural Net-
works. Cambridge, MA, USA: MIT Press, 1998.
[7] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan,
V. Vanhoucke, and A. Rabinovich, “Going deeper with convolutions,”

BIBLIOGRAPHY 98
2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR),

pp. 1–9, June 2015.
[8] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna, “Rethinking

the inception architecture for computer vision,” 2016 IEEE Conference
on Computer Vision and Pattern Recognition (CVPR), pp. 2818–2826, June

2016.
[9] O. Ronneberger, P.Fischer, and T. Brox, “U-net: Convolutional net-

works for biomedical image segmentation,” in Medical Image Computing
and Computer-Assisted Intervention – MICCAI 2015, vol. 9351 of LNCS,
(Cham), pp. 234–241, Springer International Publishing, 2015.
[10] K. Suzuki, “Pixel-based machine learning in medical imaging,” Interna-

tional Journal of Biomedical Imaging, vol. 2012, pp. 1–1, February 2012.
[11] S. Jaeger, S. Candemir, S. Antani, Y.-X. Wáng, P.-X. Lu, and G. Thoma,
“Two public chest x-ray datasets for computer-aided screening of pul-
monary diseases,” Quantitative Imaging in Medicine and Surgery, vol. 4,

pp. 475–7, December 2014.
[12] F. Pasa, V. Golkov, F. Pfeiffer, D. Cremers, and D. Pfeiffer, “Efficient deep
network architectures for fast chest x-ray tuberculosis screening and vi-
sualization,” Scientific Reports, vol. 9, pp. 1–9, December 2019.
[13] S. Jaeger, A. Karargyris, S. Candemir, L. Folio, J. Siegelman,
F. Callaghan, Z. Xue, K. Palaniappan, R. K. Singh, S. Antani, G. Thoma,

Y. Wang, P. Lu, and C. J. McDonald, “Automatic tuberculosis screening
using chest radiographs,” IEEE Transactions on Medical Imaging, vol. 33,

pp. 233–245, February 2014.
BIBLIOGRAPHY 99
[14] T. Ishigaki, S. Sakuma, and M. Ikeda, “One-shot dual-energy subtraction

chest imaging with computed radiography: Clinical evaluation of film
images,” Radiology, vol. 168, pp. 67–72, August 1988.
[15] A. Karargyris, J. Siegelman, D. Tzortzis, S. Jaeger, S. Candemir, Z. Xue,
K. Santosh, S. Vajda, S. Antani, L. Folio, and G. R Thoma, “Combination

of texture and shape features to detect pulmonary abnormalities in dig-
ital chest x-rays,” International Journal of Computer Assisted Radiology and

Surgery, vol. 11, pp. 563–576, June 2015.
[16] C. S. Pereira, H. Fernandes, A. M. Mendonça, and A. Campilho, “Detec-
tion of lung nodule candidates in chest radiographs,” in Pattern Recog-
nition and Image Analysis (J. Martí, J. M. Benedí, A. M. Mendonça, and
J. Serrat, eds.), (Berlin, Heidelberg), pp. 170–177, Springer Berlin Hei-
delberg, 2007.
[17] G. Coppini, S. Diciotti, M. Falchini, N. Villari, and G. Valli, “Neural net-
works for computer-aided diagnosis: Detection of lung nodules in chest

radiograms,” IEEE Transactions on Information Technology in Biomedicine,
vol. 7, pp. 344–57, January 2004.
[18] R. Hardie, S. Rogers, T. Wilson, and A. Rogers, “Performance analysis
of a new computer aided detection system for identifying lung nodules

on chest radiographs,” Medical Image Analysis, vol. 12, pp. 240–58, July
2008.
[19] S. Chen, K. Suzuki, and H. Macmahon, “Development and evaluation of

a computer-aided diagnostic scheme for lung nodule detection in chest
radiographs by means of two-stage nodule enhancement with support
vector classification,” Medical Physics, vol. 38, pp. 1844–58, April 2011.
BIBLIOGRAPHY 100
[20] K. Suzuki, I. Horiba, N. Sugie, and M. Nanki, “Extraction of left ven-

tricular contours from left ventriculograms by means of a neural edge
detector,” IEEE Trans. Med. Imaging, vol. 23, no. 3, pp. 330–339, 2004.
[21] K. Suzuki, I. Horiba, and N. Sugie, “Neural edge enhancer for super-
vised edge enhancement from noisy images,” IEEE Transactions on Pat-

tern Analysis and Machine Intelligence, vol. 25, pp. 1582–1596, December
2003.
[22] K. Suzuki, “Neural filter with selection of input features and its applica-
tion to image quality improvement of medical image sequences,” IEICE
Transactions on Information and Systems, vol. E85-D, pp. 1710–1718, 2002.
[23] K. Suzuki, I. Horiba, and N. Sugie, “Efficient approximation of neural

filters for removing quantum noise from images,” IEEE Transactions on
Signal Processing, vol. 50, pp. 1787–1799, July 2002.
[24] B. Sahiner, Heang-Ping Chan, N. Petrick, Datong Wei, M. A. Helvie,
D. D. Adler, and M. M. Goodsitt, “Classification of mass and normal

breast tissue: a convolution neural network classifier with spatial do-
main and texture images,” IEEE Transactions on Medical Imaging, vol. 15,
pp. 598–610, October 1996.
[25] S.-C. B. Lo, H. Li, Y. J. Wang, L. Kinnard, and M. T. Freedman, “A mul-

tiple circular path convolution neural network system for detection of
mammographic masses,” IEEE Transactions on Medical Imaging, vol. 21,
pp. 150–158, 2002.
[26] S.-C. Lo, S.-L. Lou, J.-S. Lin, M. T. Freedman, M. V. Chien, and S. Mun,
“Artificial convolution neural network techniques and applications for
lung nodule detection,” IEEE Transactions on Medical Imaging, vol. 14,

pp. 711 – 718, December 1995.
BIBLIOGRAPHY 101
[27] S.-C. B. Lo, H.-P. Chan, J.-S. Lin, H. Li, M. T. Freedman, and S. K. Mun,
“Artificial convolution neural network for medical image pattern recog-
nition,” Neural Networks, vol. 8, pp. 1201–1214, December 1995.
[28] C. Nebauer, “Evaluation of convolutional neural networks for visual
recognition,” IEEE Transactions on Neural Networks, vol. 9, pp. 685–696,

July 1998.
[29] S. Lawrence, C. Giles, A. C. Tsoi, and A. Back, “Face recognition: a con-

volutional neural-network approach,” IEEE Transactions on Neural Net-
works, vol. 8, pp. 98–113, January 1997.
[30] F. Li, H. Arimura, K. Suzuki, J. Shiraishi, Q. H. Li, H. Abe, R. M. Engel-
mann, S. Sone, H. MacMahon, and K. Doi, “Computer-aided detection

of peripheral lung cancers missed at ct: Roc analyses without and with
localization.,” Radiology, vol. 237 2, pp. 684–90, 2005.
[31] M. Loog, B. van Ginneken, and A. Schilham, “Filter learning: Applica-
tion to suppression of bony structures from chest radiographs,” Medical

Image Analysis, vol. 10, pp. 826–40, January 2007.
[32] M. Loog and B. Ginneken, “Segmentation of the posterior ribs in chest
radiographs using iterated contextual pixel classification,” IEEE Trans-
actions on Medical Imaging, vol. 25, pp. 602–611, May 2006.
[33] J.-S. Lin, S.-C. B. Lo, A. Hasegawa, M. T. Freedman, and S. K. Mun,
“Reduction of false positives in lung nodule detection using a two-

level neural classification,” IEEE Transactions on Medical Imaging, vol. 15,
pp. 206–17, 1996.

BIBLIOGRAPHY 102
[34] W. Zhang, K. Doi, M. L. Giger, R. M. Nishikawa, and R. A. Schmidt, “An

improved shift-invariant artificial neural network for computerized de-
tection of clustered microcalcifications in digital mammograms,” Medi-

cal Physics, vol. 23, no. 4, pp. 595–601, 1996.
[35] S. Oda, K. Awai, K. Suzuki, Y. Yanaga, Y. Funama, H. Macmahon, and

Y. Yamashita, “Performance of radiologists in detection of small pul-
monary nodules on chest radiographs: Effect of rib suppression with

a massive-training artificial neural network,” AJR. American Journal of
Roentgenology, vol. 193, pp. W397–402, November 2009.
[36] U. D. of Health and H. Services, “Pneumonia can be prevented - vac-
cines can help.” https://www.cdc.gov/features/pneumonia/index.

html.
[37] G. Ramalho and L. Bezerra, “Lung disease detection using feature ex-
traction and extreme learning machine,” Revista Brasileira de Engenharia
Biomedica, vol. 30, pp. 207 – 214, September 2014.
[38] F. J. Aherne, N. A. Thacker, and P. I. Rockett, “The bhattacharyya met-

ric as an absolute similarity measure for frequency coded data,” Kyber-
netika, vol. 34, pp. 363–368, 1998.
[39] G. Beylkin, “Discrete radon transform,” IEEE Transactions on Acoustics,

Speech, and Signal Processing, vol. 35, pp. 162–172, February 1987.
[40] C. Liu, J. Yuen, and A. Torralba, “Sift flow: Dense correspondence across
scenes and its applications,” IEEE Transactions on Pattern Analysis and
Machine Intelligence, vol. 33, pp. 978–994, May 2011.
[41] T. Evgeniou and M. Pontil, “Support vector machines: Theory and

applications,” Machine Learning and Its Applications: Advanced Lectures,
vol. 2049, pp. 249–257, January 2001.

BIBLIOGRAPHY 103
[42] G. Kostopoulos, I. E. Livieris, S. B. Kotsiantis, and V. Tampakas, “Cst-

voting: A semi-supervised ensemble method for classification prob-
lems,” Journal of Intelligent and Fuzzy Systems, vol. 35, pp. 99–109, 2018.
[43] K. Simonyan and A. Zisserman, “Very deep convolutional networks
for large-scale image recognition,” Computing Research Repository,

vol. abs/1409.1556, 2014.
[44] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification

with deep convolutional neural networks,” Advances in Neural Informa-
tion Processing Systems 25, pp. 1097–1105, 2012.
[45] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image
recognition,” 2016 IEEE Conference on Computer Vision and Pattern Recog-

nition (CVPR), pp. 770–778, 2016.
[46] Y. Gordienko, P. Gang, J. Hui, W. Zeng, Y. Kochura, O. Alienin,
O. Rokovyi, and S. Stirenko, “Deep learning with lung segmentation
and bone shadow exclusion techniques for chest x-ray analysis of lung
cancer,” Computing Research Repository, vol. abs/1712.07632, 2017.
[47] S. Candemir, S. Jaeger, K. Palaniappan, J. P. Musco, R. K. Singh, Z. Xue,
A. Karargyris, S. Antani, G. Thoma, and C. J. McDonald, “Lung seg-
mentation in chest radiographs using anatomical atlases with nonrigid

registration,” IEEE Transactions on Medical Imaging, vol. 33, pp. 577–590,
February 2014.
[48] K. Z. D. Kermany and M. Goldbaum, “Large dataset of labeled opti-
cal coherence tomography (oct) and chest x-ray images,” Cell, pp. 1122–
1131, 2018.
BIBLIOGRAPHY 104
[49] M. Hossin and S. M.N, “A review on evaluation metrics for data classi-
fication evaluations,” International Journal of Data Mining and Knowledge
Management Process, vol. 5, pp. 01–11, March 2015.
[50] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet:
A large-scale hierarchical image database,” IEEE Conference on Computer

Vision and Pattern Recognition (CVPR), pp. 248–255, 2009.
[51] L. Perez and J. Wang, “The effectiveness of data augmentation in image

classification using deep learning,” ArXiv, vol. abs/1712.04621, 2017.
[52] R. Wu, S. Yan, Y. Shan, Q. Dang, and G. Sun, “Deep image: Scaling up
image recognition,” ArXiv, vol. abs/1501.02876, January 2015.
[53] D. Hana, Q. Liu, and W. Fan, “A new image classification method using
cnn transfer learning and web data augmentation,” Expert Systems with
Applications, vol. 95, p. 57–71, November 2017.
[54] Q. Xu, Y.-Z. Liang, and Y.-P. Du, “Monte carlo cross-validation for se-
lecting a model and estimating the prediction error in multivariate cali-
bration,” Journal of Chemometrics, vol. 18, pp. 112 – 120, February 2004.
[55] D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,”
Computing Research Repository, vol. abs/1412.6980, 2015.
[56] Q. Kang, Q. Lao, and T. Fevens, “Nuclei segmentation in histopatho-

logical images using two-stage learning,” Medical Image Computing and
Computer Assisted Intervention (MICCAI 2019), October 2019. Accepted

June 4, 2019.
[57] M. T. Islam, M. A. Aowal, A. T. Minhaz, and K. Ashraf, “Abnormal-
ity detection and localization in chest x-rays using deep convolutional

neural networks,” ArXiv, vol. abs/1705.09850, 2017.
Bibliography 105
[58] R. H. Anuj Rohilla and A. Mittal, “Tb detection in chest radiograph us-
ing deep learning architecture,” International Journal of Advance Research
in Science and Engineering, vol. 6, pp. 1073–1084, August 2017.
[59] J. J. M. H.-J. K. Sangheum Hwang, Hyo-Eun Kim, “A novel approach for
tuberculosis screening based on deep convolutional neural networks,”

Medical Imaging 2016: Computer-Aided Diagnosis, vol. 9785, pp. 1–23,
2016.
[60] X. Wang, Y. Peng, L. Lu, Z. Lu, M. Bagheri, and R. M. Summers, “Chestx-

ray8: Hospital-scale chest x-ray database and benchmarks on weakly-
supervised classification and localization of common thorax diseases,”
2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR),
pp. 3462–3471, 2017.

Classification of Lung Diseases Using Deep Learning Models

Uploaded by

Copyright:

Available Formats

Classification of Lung Diseases Using Deep Learning Models

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Classification of Lung Diseases Using Deep Learning Models

Uploaded by

Copyright:

Available Formats

Classification of Lung Diseases Using Deep Learning

Computer Science and Software Engineering

Presented in Partial Fulfillment of the Requirements

for the Degree of

Master of Computer Science at

Montréal, Québec, Canada

c Matthew Zak, 2019

This is to certify that the thesis prepared

By: Matthew Zak

Entitled: Classification of Lung Diseases Using Deep Learning Models

and submitted in partial fulfillment of the requirements for the degree of

Master of Computer Science

originality and quality.

Signed by the Final Examining Committee:

Classification of lung diseases using deep learning models

Although deep learning-based models show high performance in the medi-

tection of patient privacy and lack of publically available medical databases.

In this thesis, we address the problem of medical data scarcity by consid-

small volume datasets (<103 samples). We implement three deep convolu-

transfer learning approach. We created a pipeline that applied segmentation

pre-trained models and simple classifiers such as shallow neural networks

We validated our techniques on the publicly available Shenzhen and Mont-

gomery datasets and compared them to the currently available solutions.

forming models trained on the Montgomery dataset however, the advantage

on the Shenzhen dataset but as previously, it is computationally less expen-

superior to your former self.”

1.1 Project objectives . . . . . . . . . . . . . . . . . . . . . . . . . . 1

2.1 Pixel/Voxel-Based Machine Learning . . . . . . . . . . . . . . 4

2.1.2 Bone Separation from Soft Tissue in Chest Radiographs

2.2.1 Extreme learning machine . . . . . . . . . . . . . . . . 8

2.2.2 Automatic CXR screening system . . . . . . . . . . . . 10

2.3 Automatic diagnostic using Deep Learning in MODS digital

3 Transfer Learning in Lung Diseases Classification 30

3.2.1 VGG16 Architecture . . . . . . . . . . . . . . . . . . . . 32

3.3.1 Residual Bloc . . . . . . . . . . . . . . . . . . . . . . . . 35

3.3.2 ResNet50 Architecture . . . . . . . . . . . . . . . . . . . 36

3.4.1 Factorizing Convolutions . . . . . . . . . . . . . . . . . 40

3.4.3 Auxiliary Classifiers . . . . . . . . . . . . . . . . . . . . 42

3.4.4 Effective Grid Size Reduction . . . . . . . . . . . . . . . 43

3.6.2 ResNet results . . . . . . . . . . . . . . . . . . . . . . . . 53

3.6.4 Results comparison . . . . . . . . . . . . . . . . . . . . . 63

4 Transfer Learning Models Accuracy Improvement in Lung Diseases

Classification Using Segmentated X-Ray Images 66

4.2 Lungs Segmentation . . . . . . . . . . . . . . . . . . . . . . . . 69

4.4.1 VGG results . . . . . . . . . . . . . . . . . . . . . . . . . 74

4.4.2 ResNet results . . . . . . . . . . . . . . . . . . . . . . . . 79

4.4.3 Inception results . . . . . . . . . . . . . . . . . . . . . . 84

4.5 Comparison of results . . . . . . . . . . . . . . . . . . . . . . . 90

5 Conclusions and Future Work 94

5.3 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

2.1 Ribs separation training samples. Image A is the input and B

2.2 Performance of a trained model in ribs removal task. Image A

2.3 Overview of the screening system developed for tuberculosis

2.4 Overview of the CST-Voting algorithm [2] . . . . . . . . . . . . 14

2.5 CST-Voting algorithm [2] . . . . . . . . . . . . . . . . . . . . . . 14

box. The right part is a positive tuberculosis culture [3] . . . . 16

224 x 224 grayscale image of a MODSM. tuberculosis culture.

The image is passed through the network, and the output of

the second fully-connected layer is a probability distribution