International Journal of Information Technology (IJIT) – Volume 7 Issue 2, Mar - Apr 2021

RESEARCH ARTICLE OPEN ACCESS

Chest Disease Detection through X-Ray using

Machine Learning

Manish Gupta [1], Rachel Calvin [2], Bhavika Desai [3], Prof. Suvarna Aranjo [4]
[1],[2],[3],[4]
Dept. of Information Technology Xavier’s Institute of Engineering, Mumbai Maharashtra - India

ABSTRACT
Chest disease have majorly affected man lives around the world.There are many chest disease such as pneumonia, asthma,
tuberculosis and many more lung diseases. If these diseases are not diagnosed in time they can turn deadly.Chest radiography
(chest X-ray)is a effective way of recognizing and the problem and cost effective at he same time. But due to lack of
professional radiologist the application of the method hasn’t reached it peak. In this paper, we will explore the possibility of
designing a computer aided diagnosis for chest X-rays using deep convolutional neural networks. Using a real-world dataset of
118055 chest X-rays with natural language diagnosis reports, we can train a multi-class classification model from images and
preform accurate diagnosis, without any prior domain knowledge. Chest X-ray (CXR)is commonly used for the diagnoses of
such lung diseases. Computeraided diagnosis (CAD)was developed for the radiologist to achieve the desired result in short
period of time.
Keyboard: - CAD, CXR, CNN, multi-label classification, problem transformation method, deep learning, image classification,
image feature extraction.

I. INTRODUCTION
classification which is called as the Convolutional neural
Medical X-rays are used to diagnose several disease in a short network in deep learning.This model provides the
period of time.Medical professionals use this technique to highest accuracy after training the dataset. This model
identify different fractures and abnormalities in different ares was introduced in 90’s for human visual perception of
of the body.This is because X rays are very effective and does recognizing things. The best known architecture in
not cause any harm to the body in any ways.Chest diseases Convolutional neural net is the LeNet architecture that
can be shown in CXR images in the form of cavitation, was used to analyze the zip codes, digits, etc. [3]. But
consolidations, infiltrates, and small broadly distributed due to it property of being data hungry researchers had to
nodules. By analyzing the X-ray’s the doctors can diagnose introduce more generalized methods in deep learning
different diseases such as effusion, pneumonia, bronchitis, [4].In multiclassification instances are introduced for the
infiltration, nodule, cardiomegaly, pneumothorax, fractures, training of the dataset by transforming problem into
and many others. In this work, we diagnose the lung diseases more single labels. The given paper introduces a model
at a much quicker pace than the radiologist. Chest X-Ray, is which will use multiclassification and deep learning
one of the most common types of radiology examination for technique for the detection of different chest
the diagnosis of lung diseases. diseases.The Ian concept is to change the multi-
However, radiologist involve the decision under uncertainty classification design into a single label concept as
Therefore, a clear output cannot be taken out[1]. Therefore, mentioned below. The following is achieved using a
Computer- Aided Diagnosis was developed to get the result publicly available dataset called NIH-Chest X-ray.
effectively with easy and in short amount of time.CAD
systems are not here to replace doctors rather help them get a II. KEY FEATURES
second opinion for the overall diagnoses. Over the past few • Recognize the disease of the chest by feeding
years we have been working on use of Computer-Aided the images to the machine learning model.
Diagnosis and Artificial Intelligence for classification of the • Machine learning model is flexible and can be
image and through the classification acquiring the accuracy . trained for any type of diseases and different
The first step towards classification is to extract the features kinds of chest diseases.
from the images which in turn will act as input to the second • Real-time images are provided to the app which
step for training[2]. The accuracy completely speedy on the in turn is integrated with Machine learning
training of the dataset.Therefore, we are using the best model model and output disease is predicted.
for the

ISSN: 2454-5414 www.ijitjournal.org Page 46

 International Journal of Information Technology (IJIT) – Volume 7 Issue 2, Mar - Apr 2021

• Information about each chest disease is available on 2019 I. Allaouzi, M. B. Ahmed: Novel Approach for
the app such as different term description related to Multi- Label CXR Classification of Common Thorax
chest or lung diseases. Diseases Problem Transformation Methods:
• Model which is built for detecting disease can be The main idea is to fit data to an algorithm by
integrated with any app because the model is converting the multi-label classification problem
converted to API and requests can be made to the into one or more binary/multi-class sub- problems, and
endpoint URL of the API. then combine their results to form the multi-label
• I prediction. Representative algorithms include Binary
t is hosted globally so anyone can access the API and Relevance [8], Random k-Labelsets [9] and Classifier
integrate it with the app. Chains [10].
Problem Adaptation Methods:
III. RELATED WORK The idea is to fit the well know techniques with the
multi-label classification model. The algorithms include
Using the publicly available dataset called ChestXray14 lazy learning techniques ML-kNN [11], an adaptation of
dataset many models have been introduced[19].In that we decision tree techniques ML-C4.5 [12], an adaptation of
have used three different CNN models and classified there kernel techniques RankSVM [13]. A novel approach
uses.[20]Present a two-staged model with recurrent neural called Ensemble methods [14] was developed on top of
network acting as a decoder. While [21]analyses is done these two approaches. It converts the problem of multi-
about which function is more suitable for the working of label classification into an ensemble of multi-label sub-
CNN model. The feasible works of ChesXNet [22] that tells problems. Representative algorithms include the
us about DenseNet-121 on the chest X-ray images, which has Random klabELsets method (RAkEL) [15], Ensemble
a modified last fully- connected layer and [23] that proposes Classifier chains (ECC) [16], and label space partitioning
a guided two-branch convolutional neural network for chest classifiers [17]. In order to find the most suitable
disease classification. The model consist of both the global approach for our case, we tried to make a comparison
and publicly evaluated model for the training and gives us the between the three above methods.
best result possible.

IV. DEEP LEARNING WITH MEDICAL VI. THE PROPOSED APPROACH

IMAGE The idea behind our approach is to combine the
effectiveness of CNN for image features extraction
Deep learning’s basically introduced to get the images which from a small image dataset and the power of the
have scare and recognize them.It basically uses the a large problem transformation methods in the task of multi-
dataset to get the value and extract the desired knowledge label classification. As shown in Figure 2, the
require for the training and testing of the data. development of the proposed method consists of four
CNN AS FEATURES EXTRACTOR parts: data description and exploration, data pre-
In order to get the fully featured network we need to processing, feature extraction part, and classification
get rid of the latest fully trained CNN model. We first have to part.
pre-trained the whole model gained and start with the training
and testing of the model.This will help us extract the right A. D A T A DESCRIPTION AND
features and accuracy in no time.The huge data should be EXPLORATION
trained and testing in such a way the the accuracy of the
1) NIH CHESTX-RAY DATASET
model should not be hindered.To get the good CNN model
we had to try and test various training models including The dataset contains 112,120 frontal CXRs from 30,805
ResNet [5], VGG-Net [6], and DenseNet [7]. As a result, we unique patients. The images are in PNG format and
chose the DenseNet-121 model which achieved the state- have a size of 1024 × 1024. CXRs are labeled with 14
ofthe-art results. common chest diseases including Atelectasis,
V. MULTI-LABEL CLASSIFICATION (MLC) Consolidation, Infiltration, Pneumothorax, Edema,
Emphysema, Fibrosis, Effusion, Pneumonia,
Multi-label image classification has gained a lot of attention Pleuralthickening, Cardiomegaly, Nodule, Mass and
in the computer vision field and has helped untacking many Hernia. If the diseases given above are on present in the
problem related to images. The binary/multi-class has one CXR, then it will be labeled as ‘‘No finding’’.
label of each image but a multi-label can have multiple image Exploring the image visually helps us gain a lot of
labels. There are different approaches have been proposed to information on the size of the image and different
address the problem of multi-label classification; they are faction for formatting it in the way possible
mainly arranged into three categories: 64280 VOLUME 7,

ISSN: 2454-5414 www.ijitjournal.org Page 47

 International Journal of Information Technology (IJIT) – Volume 7 Issue 2, Mar - Apr 2021

B. DATA PRE-PROCESSING

Data pre-processing is meant resize the

images in the way required for the training of the
model.This will help us to get the model working
for the extraction which is the next stage of the
process.With this we normalize the datable
subtracting the images as required.

C. FEATURE EXTRACTION
The important part of the CXR is to get the features that
will help us classify the images into one or multiple
possible classes. We have used the DenseNet-121 model
is used as a feature extractor. The Dense Convolutional
Network (DenseNet) [8] is a new CNN architecture that
is highly competitive object recognition benchmark
tasks. The core idea of DenseNet is to connecting all the Figure 1: DenseNet with 5
layer and letting multiple information flow through it that
match with it. As shown in Figure 2, this introduces
L×(L+1) 2 connections in an L-layer network, instead of
just L, as in traditional architectures. A DenseNet is
followed by transition layers. Each unit packs two
convolutions, each preceded by Batch TABLE 1. The
DenseNet-121 architecture.Fixed unit vector is been
generated durning the process. The layers between these
dense blocks are transition layers which perform down-
sampling of the layers features passing the network. A
detailed explanation of DenseNet-121 architecture, the
DenseNet we used in this work, is shown in Table 1.
Motivated by the results obtained by DenseNet-121 on
ChestX-ray14 dataset [18], [19], we have trained the
DenseNet-121 model on our dataset, using initial weights Figure 2: Multi-label classification pipeline
obtained from the pre- trained network, on ImageNet,
which gives a good starting point against random VII. CONCLUSION
initialization of the weights.
In this paper, we propose a new approach that
combines the effectiveness of CNN for image feature
extraction and the power of supervised multi-label
classifiers in order to tackle the task of thorax diseases
detection on CXRs. The task has been carried out with a
pre-trained DenseNet-121 model as feature extractor and
different problem transformation methods. The
evaluation process was conducted using performance
metrics average AUC. The results showed that our
method achieved great results and outperformed current
state-of-the-art on ChestX-ray14 dataset. To further
substantiate the results of this study, several
improvements could be made, such as the use of an
attention mechanism to improve CNN’s work and train
our classifier on a more balanced data set to avoid the
problem of imbalance label distribution.

REFERENCES
1. A. Brady, R. Ó. Laoide, P. McCarthy, and R.
McDermott, ‘‘Discrepancy and error in radiology:
Concepts, causes and consequences,’’ Ulster Med. J.,
TABLE 1: The DenseNet-121 Architecture vol. 81, no. 1, pp. 3–9, Jan. 2012. [Online]. Available:

ISSN: 2454-5414 www.ijitjournal.org Page 48

 International Journal of Information Technology (IJIT) – Volume 7 Issue 2, Mar - Apr 2021

https://www.ncbi.nlm.nih.gov/pmc/ articles/PMC3 609674/ 17. P. Szymański, T. Kajdanowicz, and K. Kersting,

2. W. Rawat and Z. Wang, ‘‘Deep convolutional neural ‘‘How is a data-driven approach better than random
networks for image classification: A comprehensive review,’’ choice in label space division for multi-label
Neural Comput., vol. 29, no. 9, pp. 2352–2449, Sep. 2017. classification?’’ Entropy, vol. 18, no. 8, p. 282, Jun.
[Online]. Available: https://www. mitpressjournals.org/ 2016. doi: 10.3390/e18080282.
doi/10.1162/neco_a_00990 3. Y. Lecun, L. Bottou, Y. Bengio, 18. Q. Guan, Y. Huang, Z. Zhong, Z. Zheng, L. Zheng,
and P. Haffner, ‘‘Gradient-based learning applied to document and Y. Yang. (2018). ‘‘Diagnose like a radiologist:
recognition,’’ Proc. IEEE, vol. 86, no. Attention guided convolutional neural network for
11, pp. 2278–2324, Nov. 1998. thorax disease classification.’’ [Online]. Available:
4. L. Torrey and J. Shavlik, ‘‘Transfer learning,’’ in https:// arxiv.org/abs/1801.09927.
Handbook of Research on Machine Learning Applications and J. Rubin, D. Sanghavi, C. Zhao, K. Lee, A. Qadir, and
Trends: Algorithms, Methods, and Techniques, E. Soria, J. M. Xu-Wilson. (2018). ‘‘Large scale automated reading
Martin, R. of frontal and lateral chest X- rays using dual
Magdalena, M. Martinez, and A. Serrano, Eds. convolutional neural networks.’’ [Online]. Available:
Hershey, PA, USA: IGI Global, 2009, ch. 11, p. 23. 5. K. He, https://arxiv.org/ abs/1804.07839
X. Zhang, S. Ren, and J. Sun, ‘‘Deep residual learning for
image recognition,’’ in Proc. IEEE Conf. Comput. Vis. Pattern
Recognit., Jun. 2016, pp. 770–778.
6. K. Simonyan and A. Zisserman, ‘‘Very deep convolutional
networks for large-scale image recognition,’’ in Proc. ICLR,
2015, pp. 1–14. [Online]. Available:
https://arxiv.org/abs/1409.1556 [7]G. Huang, Z. Liu, K. Q.
Weinberger, and L. Van der Maaten, ‘‘Densely connected
convolutional networks,’’ in Proc. IEEE Conf. Comput. Vis.
Pattern Recognit. (CVPR), Jul. 2016, pp. 4700–4708.
[Online]. Available: https://arxiv.org/abs/1608.06993
8. M. R. Boutell, J. Luo, X. Shen, and C. M. Brown,
‘‘Learning multi-label scene classification,’’ Pattern
Recognit., vol. 37, no. 9, pp. 1757–1771, Sep. 2004. [Online].
Available: https://www.sciencedirect.com/
science/article/pii/S0031320304001074 9. G. Tsoumakas and
I. Vlahavas, ‘‘Random k- labelsets: An ensemble method for
multilabel classification,’’ in Proc. 18th Eur. Conf. Mach.
Learn., 2007, pp. 406–417.
10. J. Read, B. Pfahringer, G. Holmes, and E. Frank,
‘‘Classifier chains for multi-label classification,’’ in Machine
Learning and Knowledge Discovery in Databases, vol. 5782.
2009, pp. 254–269. doi:
10.1007/s10994-011- 5256-5.
11. M.-L. Zhang and Z.-H. Zhou, ‘‘A k-nearest neighbor based
algorithm for multi-label
classification,’’ in Proc. IEEE Int. Conf. Granular Comput.
(GrC), Jul. 2005, pp. 718–721. [12] A. Clare and R. D. King,
‘‘Knowledge discovery in multi-label phenotype data,’’ in
Proc. 5th Eur. Conf. PKDD, 2001, pp. 42–53.
13. A. Jiang, C. Wang, and Y. Zhu, ‘‘Calibrated rank- SVM
for multi-label image categorization,’’ in Proc.
IEEE Int. Joint Conf. Neural Netw. (IJCNN), Jun. 2008, pp.
1450–1455.
14. G. Madjarov, D. Kocev, D. Gjorgjevikj, and S. Džeroski,
‘‘An extensive experimental comparison of methods for multi-
label learning,’’ Pattern Recognit., vol. 45, no. 9, pp. 3084–
3104, Sep. 2012. [Online]. Available:
https://www.sciencedirect.com/science/ article/pii/S003
1320312001203
15. G. Tsoumakas, I. Katakis, and I. Vlahavas, ‘‘Random k-
labelsets for multilabel classification,’’ IEEE Trans. Knowl.
Data Eng., vol. 23, no. 7, pp. 1079–1089, Jul. 2011.
16. J. Read, ‘‘A pruned problem transformation method for
multi-label classification,’’ in Proc. New Zealand Comput.
Sci. Res. Student Conf., 2008, pp. 143–150.

ISSN: 2454-5414 www.ijitjournal.org Page 49