Simplifying process of computer aided diagnosis Essay

Simplifying process of Computer Aided Diagnosis

of Tuberculosis

Parth Patel

Department of Computer Science and


SRM Institute of Science and



([email protected])

S. Usha Kiruthika

Department of Computer Science and


SRM Institute of Science and



[email protected]

Abstract— Chest x-rays are the most commonly performed

which are costly diagnostic imaging tests ordered by

physicians. Especially in countries like India where the ratio

of radiologist or experts per capita are low in numbers and

the high quality resources are many patients remain un-

diagnosed or diagnosed lately due to lack of machineries.

The aim of this study is to develop and validate a deep

learning system to detect chest x-ray abnormalities and

hence detect Tuberculosis (TB) and to provide a tool for

Computer Aided Diagnosis (CAD).

by exploring existing systems of Image Processing and

Deep learning architectures like U-net, Faster RCNNs",The

study is aimed to achieve radiologist level detection as well

as lower False Negative detection of TB by using ensemble

as well as locally curated data-set and algorithms. Also, I am

I'm trying to simplify the process of detection by developing

a web API which takes an X-ray as an Input and classify as

potential TB cases instantly.

Keywords :- Tuberculosis, U-net, Chest X-rays, Computer

Aided Diagnosis(CAD), Radiology, Deep Learning


TB is one of the top ten causes of death world wide.In 2017",

10.6 million cases of TB were reported. Among those 10

million cases 1.7 million people met death due to the disease.

In populous countries like India, there is a shortage of

radiologists, Over 95percent of cases and deaths are in

developing countries like India. India has a severely

imbalanced ratio of 1:100",000 between radiologist and

people. Interpreting chest x-rays is one of the most complex

tasks for radiologists. So there is an emerging need for

computer-aided diagnosis popularly known as (CAD) which

can assist Doctors and Practitioners in diagnosing disease on

early stage and reduce the false negative error rate. The CAD

can also reduce the cost associated with the Diagnosis which

can help schemes like ’Ayushyaman Bharat’to meet success.


Semantic pixel-wise segmentation is an active topic of

research, fuelled by challenging datasets . Before the

arrival of deep networks, the best-performing methods

mostly relied on hand engineered features classifying

pixels independently. Typically, a patch is fed into a

classifier e.g. Random Forest or Boosting to predict the

class probabilities of the center pixel. Features based on

appearance for SfM and appearance have been explored for

the CamVid road scene understanding test . These perpixel

noisy predictions (often called unary terms) from the

classifiers are then smoothed by using a pair-wise or higher

order CRF to improve the accuracy.


We know that the labeled medical images are very less in

numbers and hence there is a strong need for data

augmentation. There is also need for faster and efficient

segmentation. So I propose the system architecture of CNN",

called U-Net which can perform both augmentation as well

as segmentation in fast and efficient way. Also it can output

high resolution image from low resolution inputs. The U-

Net architecture is built upon the Fully Convolutional

Network and modified in a way that it yields better

segmentation in medical imaging. Compared to FCN-8, the

two main differences are (1) U-net is symmetric and (2) the

skip connections between the downsampling path and the

upsampling path apply a concatenation operator instead of a

sum. These skip connections intend to provide local

information to the global information while upsampling.

Because of its symmetry, the network has a large number of

feature maps in the upsampling path, which allows to

transfer information. By comparison, the basic FCN

architecture only had number of classes feature maps in its

upsampling path.


sampling path

The contracting path is composed of 4 blocks. Each

block is composed of :-

 3x3 Convolution Layer + activation function (with

batch normalization)

 3x3 Convolution Layer + activation function (with

batch normalization)

 2x2 Max Pooling

Note that the number of feature maps doubles at each

pooling, starting with 64 feature maps for the first block",

128 for the second, and so on. The purpose of this

contracting path is to capture the context of the input image

in order to be able to do segmentation. This coarse

contextual information will then be transfered to the

upsampling path by means of skip connections. This part of

the network is between the contracting and expanding

paths. The bottleneck is built from simply 2 convolutional

layers (with batch normalization), with dropout.


sampling path

The expanding path is also composed of 4 blocks. Each of

these blocks is composed of :-

 Deconvolution layer with stride 2.

 Concatenation with the corresponding cropped

feature map from the contracting path.

 3x3 Convolution layer + activation function

(with batch normalization).

 3x3 Convolution layer + activation function

(with batch normalization).

The purpose of this expanding path is to enable precise

localization combined with contextual information from the

contracting path.


For this experiment, we plan to use three types of

dataset. One is Large dataset made public by National

Institute of Health popularly known as NIH. NIH center has

open sourced over 100",000 anonymous chest x-ray images

with respective medical annotations and data for

researchers across the globe.

A.Shenzhen Hospital X-ray Set

X-ray images in this data set have been collected by

Shenzhen No.3 Hospital in Shenzhen, China. The CXRs

were procured as part of the regular supervision at the

Hospital. It includes CXRs in JPEG format. There are 340

normal CXRs and 275 CXRs with abnormalities showing

various indications of tuberculosis .

B. Montgomery County X-Ray


CXRs images in this data set have been acquired from

the TB curb program by Department of Health and Human

Services of Montgomery County, USA. This set contains

138 CXRs, among which 80 CXRs are normal and 58

CXRs are abnormal with manifestations of tuberculosis.

All images are de-identified and available in DICOM

format. The set covers a wide range of abnormalities",

including effusions and military patterns. The data set

includes radiology readings available as a text file Apart

from above three Datasets for CXRs we also plan to use

various augmentation methods to enhance dataset and

ensemble more publicly available datasets to enhance

accuracy of the model.


In this work, I present a system for medical application

of chest pathology detection in x-rays which uses

convolutional neural networks that are learned from a non-

medical archive (ImageNet). We show that a combination

of deep learning features to achieve better performance.

Additionally, we can also show that deep CNN layers

achieve better results compared to shallow layers. Our

results may demonstrate the feasibility of detecting

pathology in chest x-ray using deep learning approaches

based on non-medical learning. Deep learning methods

have not been tested for chest pathology detection for our

knowledge, especially not with non-medical archive

learning. This is a first-of-its-kind experiment that shows

that Deep learning with ImageNet training may be

sufficient for general medical image recognition tasks. We

can also show that using Larger dataset and more Image

features as well as medical features helps in increasing the

accuracy and reduce false negative reporting error.

References are important to the reader; therefore, each

citation must be complete and correct. If at all possible",

references should be commonly available publications.


[1] Global Tuberculosis Control: WHO Report. World Health

Organization (2018).

[2] Computer-assisted diagnosis of tuberculosis: a first-order statistical

approach to chest radiograph .

[3] Improving Tuberculosis Diagnostics using Deep Learning and Mobile

Health Technologies among Resource-poor and Marginalized


[4] SegNet: A Deep Convolutional Encoder-Decoder Architecture for

Image Segmentation.

[5] Commentary - Radiology in India: The Next Decade.




[7] Computer-Aided Diagnosis in Medical Imaging: Historical Review",

Current Status and Future Potential.

[8] Optical Character Recognition by Ravina Mithe, Supriya Indalkar",


[9] Deep Learning in Neural Networks by Jurgen Schmidhuber, 2014.

[10] Lung Image Segmentation Using Deep Learning Methods and

Convolutional Neural Networks.

How to cite this essay: