2019 IEEE Canadian Conference of Electrical and Computer Engineering (CCECE)

Automatic Detection of Pneumonia on Compressed

Sensing Images using Deep Learning
Sheikh Rafiul Islam∗ , Santi P. Maity† , Ajoy Kumar Ray‡ and Mrinal Mandal§
∗ ,†Dept. of Information Technology
Indian Institute of Engineering Science and Technology, Shibpur, INDIA-711103
‡ Dept. of Electronics and Electrical Communication Engineering

Indian Institute of Technology, Kharagpur, INDIA-721302

§ Dept. of Electrical & Computer Engineering

University of Alberta, Edmonton, Canada

Email: ∗ [email protected], † [email protected], ‡ ajoy [email protected], § [email protected]

Abstract—Pneumonia is one of the life threatening very framework that enables detection of pneumonia more accu-
common disease and needs proper diagnosis at an early stage rately.
for proper treatment of recovery. Chest X-ray is used as an Deep learning (DL) is a technique of extracting patterns
imagining modality to identify the disease by a professional
radiologist. This paper suggests a Compressed Sensing (CS) based from data through an artificial neural network (ANN) with
deep learning framework for automatic detection of pneumonia a large number of interconnected units in order to perform
on X-ray images to assist the medical practitioners. Extensive real world complex tasks especially related to images [4].
simulation results show that the proposed approach enables Sometimes, the deep learning based systems perform better
detection of pneumonia with 97.34% prediction accuracy and than human in term of accuracy in pattern recognition. The
an improvement on reconstruction quality of the X-ray images
in terms of PSNR by 1±0.76 dB and SSIM by 0.2±0.05 using the deep learning is chosen as a very natural application in the
proposed method compared to the other state-of-the-art methods. field of medical radiology as the diagnosis process as it relies
on extracting useful information from images. The use of
Index Terms—Compressed sensing, Pneumonia detection, Deep deep learning in CS paradigm ensures convergence bounds
Learning, CNN for the recovery of a signal in an underdetermined system
[5]. An underdetermined system, instead of following the
I. I NTRODUCTION Nyquist sampling theorem, enables signal acquisition process
Pneumonia is a very common infective disease caused by by reducing the required sample size with significantly higher
bacteria, viruses, or fungi [1]. In severe case, it leads to order [5]–[7]. Hence, the combining effect of DL with the
death specially in children younger than 5 years of age, older CS leads to faster signal acquisition along with a guaranteed
adults and people with other diseases or impaired immune improvement in diagnosis accuracy.
systems. In every year, around 1 million people are detected The rest of the paper is organized as follows: a brief review
with pneumonia and around 50, 000 die from the disease on state-of-the-art methods and scope of the present work
in the United States (U.S.) like country [2]. To diagnose are reported in Section II. The proposed system model using
pneumonia, chest X-ray is used by the medical practitioners as DL framework for CS image reconstruction and pneumonia
the best imaging modality. However, proper analysis of the X- detection is described in Section III. Section IV presents
ray images need a radiologist with rich domain expertise and the simulation results and performance evaluation. Finally,
experience. According to an estimate by the World Health conclusions and scope of the future works are stated in Section
Organization (WHO), two thirds of the people in the world V.
don’t have access to a radiologist for diagnosis of their disease
yet. It would be very helpful if an automated computerized II. L ITERATURE R EVIEW AND S COPE OF P RESENT W ORK
system is developed to analyze the X-ray images for assisting This section presents a brief review on the related works and
the medical advisor to diagnose appropriately. To offer proper describes the scope and contributions of the present work.
medical services in remote villages, often imaging to be done
at one end and the medical advisors are present at far end A. Literature Review
which needs transmission of images over wireless channels. Over the last decade, several machine learning based auto-
To avail and adapt the limited bandwidth of wireless channel, mated methods for identifying different types of pneumonia
far end image reconstruction (to be used for diagnosis) to be have been widely studied [8]–[11]. Fiszman et al [8] used a
accomplished from sub-sample measurements [3]. This faster natural language processing (NLP) tool to identify acute bac-
sensing i.e. sub-sample sensing is also used for reduction in terial pneumonia-related disease in chest X-ray. Performance
radiation from X-ray imaging. Recently, Compressed Sensing of this type of resource intensive application is very much
(CS) is used in various medical imaging and a deep learning comparable to that of the human expert. Chapman et al [9]

978-1-7281-0319-8/19/$31.00 ©2019 IEEE

 2019 IEEE Canadian Conference of Electrical and Computer Engineering (CCECE)

Convolutional Network
demonstrated three computerized methods using a rule base, a
probabilistic Bayesian network, and a decision tree to diagnose
the chest X-ray report associated with acute bacterial pneumo-
nia. In [10], a study of feasibility of an NLP-based monitoring
system is done to identify healthcare-associated pneumonia 32 × 32 × 128
64 × 64 × 64 x̂ 
in neonates. However, practical clinical applications of these
types of methods are limited due to the dependency on the Label

information extracted from the narrative reports of the patients. 1

0
Parveen et al [12] reports an unsupervised fuzzy c-means 8 × 8 × 1024
M 16 × 16 × 512

classification learning algorithm to detect pneumonia infected x y ∈ R

4096
16 × 16 × 128
32 × 32 × 256

X-ray images. This approach improves classification accuracy 2048 1028

as fuzzy c-means allocate weights to all the pixels of the input Convolutional Network Fully­Connected Network

X-ray images. Rajpurkar et al [11] demonstrated ChexNet, Fig. 1: Proposed CS based DL model for pneumonia detection and reconstruction of the
a 121-layer deep convolutional neural network (CNN), that X-ray images
provides the probability of detecting or identifying pneumonia
using a heatmap to localize the area of the infection. Kermany A. CS Image Reconstruction Model
et al. [13] introduced a transfer learning based DL framework The image reconstruction is considered as an inverse prob-
to diagnose pediatric pneumonia using chest X-ray images. lem which seeks an estimation x̂ of the image by minimizing
However, none of the methods are exploited to classify X-ray the following objective function
images with pneumonia for the CS framework to meet the
need of remote end analysis. x̂ = arg min k Φx − y k22 +λϕ(x) (2)
x∈S
B. Scope and Contributions of Present Work
where S is a convex set, λ > 0 trade offs the weights
An accurate reconstruction followed by classification algo-
between the terms in the objective function and ϕ is a
rithm plays a very important role in developing an application
regularization parameter.
specific medical imaging system to diagnose a disease reliably.
The objective function in Equation 2 can be equivalently
An image classification system operated under CS framework
learned by a CNN through updating weights W using proximal
needs to extract the most useful features from the observations
(or projected) gradient descent [5].
for accurate classification and also to address an ill-posed
inverse problem to reconstruct the X-ray image. Therefore, 

it is very much essential to develop a robust method which x̂n = Q R x̂n−1 ; Φ, y, % (3)
can be easily implemented in a low cost X-ray imaging
where, Q is defined as a quasi-projection operator onto the
system to detect pneumonia more accurately. To this aim, this
set of artifact-free images,
n is the iteration number and α is
paper suggests a DL based CS framework with the following
the step size. The C x̂(n) denotes the cost function  which
contributions:
• A multi-layer convolutional neural network (CNN) is
is defined as a reconstruction operator R x; Φ, y, % to learn
used to extract features from CS measurements for clas- Equation 2 by the CNN. The % is a set of parameters for the
sifying the X-ray images of pneumonia patient. reconstruction method.
• A sub-channel is used to provide the reconstructed X-ray
image which helps to verify manually whether pneumonia B. The classification Model
is present or not. A simple multi-layer network, with 4 convolutional and
• Extensive simulation results show that the proposed ap- 3 fully connected layers as hidden layers, is used to detect
proach can classify the pneumonia infected X-ray images pneumonia for the CS framework as shown in Figure 1. To
with 97.34% prediction accuracy. introduce non-linearity in the network, all the hidden layers
use leaky Rectified linear unit (ReLU) activation function that
III. P ROPOSED S YSTEM M ODEL
has a non-zero gradient over its entire domain. The leaky ReLu
This section presents the proposed system model using (LReLU) activation function is defined as
DL framework for CS image reconstruction and pneumonia (
detection. In CS, a sparse signal x ∈ <N is recovered y if y > 0
from a measurement vector y ∈ <M (where M << N ). flR (y) =
0.001y otherwise
Mathematically, the measurement vector can be represented
using the following equation As shown in Figure 1, after reshaping the input measure-
ment vector y, 2 − d convolutional followed by up-sampling
y = Φx + η (1)
operations are performed to reconstruct the X-ray image. The
where Φ ∈ <M ×N is a linear operator called as sensing matrix last three layers are added as fully connected network for
and η ∈ <N is the noise on the sensing signal. classifying the input with an appropriate label. The output
 2019 IEEE Canadian Conference of Electrical and Computer Engineering (CCECE)

layer uses a sigmoid function with 2 units (one for pneumonia low throughout the network and also adds regularization effect
and another one for normal) as defined follows: to reduce overfitting of the network on the training data [15].
1 TABLE I: Hyper Parameters used in the Training
fS (y) =
1 + e−ΘT y
Parameter Name Value
where, Θ denotes parameters of the model, defined as the
weight matrices W and bias vectors b. Initial Learning Rate, α 0.0018
Momentum:γ, τ 0,0
Now, the output of the final layer can be expressed mathe-
Dropout Rate 0.4
matically as
Optimizer used Gradient Descent
 
Layer activations Leaky ReLu
f (y; Θ) = fS W hF flR W hF −1 y + bhF −1 + bhF


(4)
Figure 2 shows the plot of the variation of the classification
The output of the final layer can be defined as categorical
accuracy evaluated for both the training and the validation
probability distribution:
data-sets against each epoch. During the learning process,
f (y)c = p̂(Lo = c|y) the training and the validation accuracy reach 98.82% and
99.80%, respectively after completion of 400 epochs. Variation
where Lo denotes the output label and c ∈ {0, 1} is one of of the MSE loss in every epoch is also studied as plotted
the 2 output classes. in Figure 3 for both the training and the validation data-
The parameters of the network is trained by minimizing the sets. The training and the validation loss go on reducing with
cross-entropy (negative log-likelihood) shown in Equation 5 the development of the learning process of the network as
between the models output and the target distribution using expected.
stochastic gradient descent (SGD) on mini-batches.
  1
X
L f (y); Lo = − log f (y)c = − log f (y)Lo (5)
c=0

IV. S IMULATION R ESULTS AND D ISCUSSION

This section illustrates performance of the proposed ap-
proach in term of prediction accuracy of the pneumonia. The
performance of the reconstruction network is also evaluated
on the reconstructed X-ray images in terms of quantitative
analysis assessment measures like Peak-Signal-to-Noise-Ratio
(PSNR) and mean Structural SIMilarity (SSIM). All the simu-
lations presented here are carried out in Google’s online cloud
support tool for DL training known as Colaboratory (Colab)
with ‘runtime type’ and ‘hardware accelerator’ chosen as
Fig. 2: Performance Evaluation of Reconstruction Network: Reconstruction Loss vs.
‘Python3’ and ‘GPU’, respectively. The KERAS 2.0.8 libraries Epoch.
and packages with TensorFlow 1.0 as backend are used to
describe the network model in the Colab.
The data-set used for this work is obtained from Kaggle [14]
which contains 5863 X-ray images (JPEG) of two categories
(Pneumonia/Normal). This work uses 70% of the data-set
for the training, 25% for the validation and 5% for the test
purpose. All the simulations consider 25% most significant
coefficients as measurements (M) captured from the discrete
cosine transform (DCT) as sparsifying basis of the images.
In the proposed approach, the measurements are projected
through the Inverse DCT (IDCT) operator and re-sized the
images to 128 × 128 for feeding the network as shown in
Figure 1. All the convolutional layers perform the convolution
operations using kernels of size 5×5×TS and stride of 1. Here,
the Ts denotes the target size of the output activation maps. Fig. 3: Performance Evaluation of Reconstruction Network: Reconstruction Loss vs.
Epoch.
The subsequent layer performs pooling/upsampling using a
kernel of size 2 × 2 with stride of 2. The hyper parameters Performance of the trained model is also assessed using
used for the training of the network are mentioned with their the test data-set and compared with the existing methods
respective values in Table I. This work uses batch normaliza- like F-cMeans [12], DL [13] and ChexNet [11], in term of
tion (BN) which ensures no activation passes too high or too the prediction accuracy as reported in Table II. A study on
 2019 IEEE Canadian Conference of Electrical and Computer Engineering (CCECE)

TABLE III: Performance comparison of reconstructed X-ray images using different

the prediction accuracy is also performed with respect to methods
the varying % of measurements input as shown in Figure 4.
Figure 5 presents the reconstructed X-ray images using the Method Used M (%) PSNR (dB) SSIM
DL-Guided CS [5], conventional CS [16] and the proposed DL-Guided CS [5] 26.22 0.78
method to compare their visual quality. The average values of Conventional CS [16] 28.77 0.88
25
the PSNR and SSIM of the reconstructed images using 25% PM 29.40 0.91
measurements of the different test cases with respect to the
observations for detecting pneumonia with desired accuracy
full scale samples are reported in Table III.
compared to the conventional method. Simulation results show
TABLE II: Prediction Accuracy Comparison
that the proposed CS based DL scheme offers very high
prediction accuracy on X-ray images and its comparable
Method Used F-cMeans [12] DL [13] ChexNet [11] PM
reconstruction quality with respect to the full scale images
Prediction
85.87 91.16 95.28 97.34 in terms of PSNR with 28.53 ± 2.07dB and SSIM with
Accuracy (% )
0.92 ± 0.04.
The proposed method may be extended as a generalized
automatic computerized system to assist medical professionals
by localizing the region of interest (ROI) like brain tumor,
cancerous cell, kidney and gallbladder stone etc.
R EFERENCES
[1] J. H. Bates, G. D. Campbell, A. L. Barton, G. A. McCracken, P. N.
Morgan, E. B. Moses, and C. M. Davis, “Microbial etiology of acute
pneumonia in hospitalized patients,” Chest, vol. 101, no. 4, pp. 1005–
1012, 1992.
[2] url:https://www.cdc.gov/features/pneumonia/index.html.
[3] S. S. Kar and S. P. Maity, “Gradation of diabetic retinopathy on
reconstructed image using compressed sensing,” IET Image Processing,
vol. 12, no. 11, pp. 1956–1963, 2018.
[4] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521,
no. 7553, p. 436, 2015.
Fig. 4: Performance Evaluation: Accuracy vs. Measurement (%) [5] B. Kelly, T. P. Matthews, and M. A. Anastasio, “Deep learning-
guided image reconstruction from incomplete data,” ArXiv preprint
arXiv:1709.00584, 2017.
[6] D. L. Donoho, “Compressed sensing,” IEEE Transactions on Informa-
tion Theory, vol. 52, no. 4, pp. 1289–1306, 2006.
[7] R. G. Baraniuk, “Compressive sensing,” IEEE Signal Processing Mag-
azine, vol. 24, no. 4, pp. 118–120, 2007.
[8] M. Fiszman, W. W. Chapman, S. R. Evans, and P. J. Haug, “Automatic
identification of pneumonia related concepts on chest x-ray reports.,” in
Proc. of the AMIA Symposium, p. 67, American Medical Informatics
Association, 1999.
[9] W. W. Chapman, M. Fizman, B. E. Chapman, and P. J. Haug, “A
comparison of classification algorithms to automatically identify chest x-
ray reports that support pneumonia,” Journal of Biomedical Informatics,
vol. 34, no. 1, pp. 4–14, 2001.
(a) (b) [10] E. A. Mendonca, J. Haas, L. Shagina, E. Larson, and C. Friedman,
“Extracting information on pneumonia in infants using natural language
processing of radiology reports,” Journal of Biomedical Iinformatics,
vol. 38, no. 4, pp. 314–321, 2005.
[11] P. Rajpurkar, J. Irvin, K. Zhu, B. Yang, H. Mehta, T. Duan, D. Ding,
A. Bagul, C. Langlotz, K. Shpanskaya, et al., “Chexnet: Radiologist-
level pneumonia detection on chest x-rays with deep learning,” ArXiv
preprint arXiv:1711.05225, 2017.
[12] N. Parveen and M. M. Sathik, “Detection of pneumonia in chest X-
ray images,” Journal of X-ray Science and Technology, vol. 19, no. 4,
pp. 423–428, 2011.
[13] D. S. Kermany, M. Goldbaum, W. Cai, C. C. Valentim, H. Liang, S. L.
Baxter, A. McKeown, G. Yang, X. Wu, F. Yan, et al., “Identifying
(c) (d) medical diagnoses and treatable diseases by image-based deep learning,”
Fig. 5: Performance comparison: (a) Full scale X-ray image (b) Reconstructed X-ray Cell, vol. 172, no. 5, pp. 1122–1131, 2018.
image using DL-Guided CS [5] (c) Reconstructed X-ray image using conventional CS [14] url:https://www.kaggle.com/paultimothymooney/chest-xray pneumonia.
[16] (d) Reconstructed X-ray image using PM [15] S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep
network training by reducing internal covariate shift,” ArXiv preprint
V. C ONCLUSIONS AND F UTURE WORKS arXiv:1502.03167, 2015.
[16] P. Boufounos, M. F. Duarte, and R. G. Baraniuk, “Sparse signal recon-
This paper proposes an automatic computerized system for struction from noisy compressive measurements using cross validation,”
detecting pneumonia using a CS based DL model. The study in 14th IEEE Workshop on Statistical Signal Processing, pp. 299–303,
2007.
shows that the use of DL in CS framework reduces the required