Proceedings of the International Conference on Artificial Intelligence and Smart Systems (ICAIS-2021)

IEEE Xplore Part Number: CFP21OAB-ART; ISBN: 978-1-7281-9537-7

Pneumonia Detection Using Image Processing And

Deep Learning
Harsh Agrawal
2021 International Conference on Artificial Intelligence and Smart Systems (ICAIS) | 978-1-7281-9537-7/20/$31.00 ©2021 IEEE | DOI: 10.1109/ICAIS50930.2021.9395895

Student
Narsee Monjee Institute of Management Studies
Mumbai, India
[email protected]

Abstract— Pneumonia is the leading cause of death all around large amount of healthcare data and training of deep learning
the world. On average, it kills 700,000 children per year on models have become very cheap. CNN are able to learn
average and affects 7% of the world's population. This condition temporal and spatial features from images easily and hence are
is diagnosed mainly using X-rays. We have devised a novel Image suited for image-based tasks. They have proven effectiveness
processing-based Deep Learning approach to detect Pneumonia
using the concept of Transfer Learning and Image
in extracting appropriate features with no human
Augmentation. We were able to achieve a recall score of 97.44% superintendence. A CNN can have multiple convolutional and
and an accuracy of 96.00%. activation layers, often mixed with pooling layers, and is
trained using backpropagation and gradient descent. The input
Keywords—Computer-Aided Diagnostics, Computer Vision, Deep image is passed through a combination of convolutional and
Learning, Image Processing, Pneumonia Detection, Transfer pooling layers, which is followed by a set of dense and dropout
learning. layers. The dropout layers are used to reduce overfitting. The
number of units in the penultimate dense layer depends on the
1. INTRODUCTION number of classes we are classifying. Recently, CNN based
Pneumonia is a bacterial infection of the lungs that can vary algorithms have been used in detecting various diseases like
in severity from mild to fatal. It infects a person when an Skin Cancer [15], Brain Tumor [16] and even Covid-19 [17]
infection causes the air sacs in their lungs to fill with fluid or and have shown very promising results.
pus, making it difficult for them to inhale enough oxygen to In most cases of Pneumonia, the bottom tips of the lungs are
keep their bloodstreams healthy. Respiratory viruses like not clearly defined in the X-ray and a large part of the lung on
influenza, rhinoviruses (common cold) are the main cause of the right side is not visible due to the infection present in the air
this disease. Almost everybody is at risk of contracting this spaces. All types of Pneumonia have been presented in Figure
lung infection, but children under the age of two and 1. One problem with this disease is that the very presence of
individuals over the age of 65 are the most vulnerable. That's this disease is frequently confused with other diseases and,
because their immune systems won't be able to combat it in thus, radiologists find it difficult to identify this disease. Deep
most cases. It is a ubiquitous disease all around the world. learning techniques have the potential to solve all these
Pneumonia constituted 15% of all the child deaths in 2017 and problems with accuracies comparable to/greater than humans.
caused over 500,000 deaths of children under 5 in India itself They have been used in past by various researchers and all
[14]. It is even more severe in case of people above 70 years of those methods are compared with the proposed solution in the
age as 261 people out of every 100,000 people died of result section.
pneumonia in 2017 in that age range [14]. The World Health
Organization says that, “Every year, it kills an estimated 1.4 This paper proposes a novel method for detecting Pneumonia
million children under the age of 5 years, accounting for 18% from X-rays using Deep Learning and Image Preprocessing.
of all deaths of children under five years old worldwide [11]. This paper strongly emphasizes the importance of Image
Pneumonia can easily go overlooked until it has reached a Processing as it can drastically improve performance in deep
point of incurability and can result in fatality. Pneumonia learning models and, hence this paper attains its novelty from
affects children and families worldwide but is most widespread the Image Processing techniques used. We have also used the
in South Asia and Sub-Saharan Africa [20]. Children can be concept of Transfer Learning [13] by using Resnet50v2 as the
protected from pneumonia by simple interventions and can be base of our custom model. We have used Morphological
treated with low-cost, low-tech medication and care”. Operators for Image Processing which increased the
performance of our model significantly. Our method can aid
In recent times, Computer-assisted diagnostics with artificial medical practitioners in making decisions on the diagnosis of
intelligence-based solutions have been highly popular [18]. the disease and can help to reduce the fatality it causes. If
Convolutional Neural Network-motivated deep learning facilitated efficiently and correctly this method can have a
algorithms have started to become the quality choice for significant impact on the lives of many people around the
medical image classifications because of the availability of a globe.

978-1-7281-9537-7/21/$31.00 ©2021 IEEE 67

Authorized licensed use limited to: International Islamic University Malaysia. Downloaded on July 20,2022 at 08:25:36 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the International Conference on Artificial Intelligence and Smart Systems (ICAIS-2021)
IEEE Xplore Part Number: CFP21OAB-ART; ISBN: 978-1-7281-9537-7

Here are the types of Pneumonia: 2. METHODOLOGY

2.1. DATASET
We have used the Pneumonia Chest X-ray Dataset provided
by GLOWM (The Global Library of Women’s Medicine) [1].
The original dataset consists of 5216 training images and 624
test images and a 16-image validation set. As the number of
images in the validation set is strangely low, we have merged
the validation set into the training set. The training dataset has
a great imbalance in classes (3883 images for Pneumonia class
and 1349 images for Normal class) so we have used the
technique of Upsampling to avoid biased results. In
Upsampling we try to remove bias in the dataset by copying a
random sample of images of weakly represented class until the
imbalance ratio is improved. Therefore, the final dataset
consists of 6588 training images (3883 images for Pneumonia
class and 2705 images for Normal class) and 624 test images
(390 images for Pneumonia class 234 images for Normal
class). As the validation set was merged, we have used the test
set as the validation set.

Figure 2 – Stucture of the dataset

2.2. BLOCK DIAGRAM OF PROPOSED SYSTEM

This is the four-stage method block diagram:

Figure 1 – Types of Pneumonia

In this paper, we start by discussing the methodology we used -

starting from the dataset to the explanation of each step in the
algorithm, through the performance metrics to the result and
conclusion
Figure 3 – Block Diagram of Proposed Work

978-1-7281-9537-7/21/$31.00 ©2021 IEEE 68

Authorized licensed use limited to: International Islamic University Malaysia. Downloaded on July 20,2022 at 08:25:36 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the International Conference on Artificial Intelligence and Smart Systems (ICAIS-2021)
IEEE Xplore Part Number: CFP21OAB-ART; ISBN: 978-1-7281-9537-7

2.3. IMAGE PREPROCESSING High Pass Filter:

This is the first step in the proposed solution. We first
converted all images to grayscale because all the image
processing techniques work best in grayscale, and then applied
Image Processing. We tested various plausible Image
Processing techniques including Filters, Edge Detection
Algorithms, Thresholding and Morphological Operators. After
rigorous testing, we narrowed down to 2 techniques -
Histogram Equalization followed by a combination of dilation
and erosion Morphological Operators. Here are all the Image
processing techniques we tested, with observation for each
method stating the reason why we chose/didn’t choose a Figure 6 – Chest X-ray images of (a) Normal and (b) Pneumonia
particular method: patient after applying High Pass Filter.

Original Images: Observation:

We can observe from the output that all the important features
that are required for Pneumonia detection are lost after
applying a high pass filter and only the rib-cage is remaining
and therefore a high pass filter cannot be used for pre-
processing.
Low Pass Filter:

Figure 4 - Chest X-ray images of (a) Normal and (b) Pneumonia

patient.
Histogram Equalization:

Figure 7 – Chest X-ray images of (a) Normal and (b) Pneumonia

patient after applying Low Pass Filter.
Observation:
From the output, we can observe that there is no significant
difference in the image after applying the low pass filter,
Figure 5 - Histogram Equalized Chest X-ray images of (a) Normal and except that it is a bit blurred which ultimately would end up
(b) Pneumonia patient. hurting the model performance. Hence, Low pass filters cannot
be used for pre-processing.
Observation:
Sobel Operator:
Histogram equalization is used to enhance the contrast of an
image, and that’s what it exactly does here. This enables areas
of lower contrast to gain a greater contrast. It accomplishes this
by spreading out the intensity values according to their
frequencies. And as the expanded intensity values seem to
highlight the essential features it would be a good fit for
preprocessing.

Figure 8 – Chest X-ray images of (a) Normal and (b) Pneumonia

patient after applying Sobel filter.
Observation:
The Sobel operator works well with medical images, as we
observed from our literature survey. However here we observe
that the edges of the ribcage are very clearly detected, which is

978-1-7281-9537-7/21/$31.00 ©2021 IEEE 69

Authorized licensed use limited to: International Islamic University Malaysia. Downloaded on July 20,2022 at 08:25:36 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the International Conference on Artificial Intelligence and Smart Systems (ICAIS-2021)
IEEE Xplore Part Number: CFP21OAB-ART; ISBN: 978-1-7281-9537-7

opposite of what we want. Hence Sobel operator cannot be Closing technique) as the pre-processing pipeline for our
used for training. training and testing images.
Thresholding: 2.4 IMAGE AUGMENTATION
Image augmentation is a technique to generate new training
data artificially from current training data. This is achieved
through the use of specific techniques to examples from the
training data that create new and distinct training examples.
Image data augmentation involves the creation of transformed
image copies in the training dataset which are of the same class
as the original image. Transformations involve a variety of
image processing operations, such as flips, shifts, zoom, and
much more. We have used only two transforms – Zooming
Figure 9 – Thresholded Chest X-ray images of (a) Normal and (b) (30%) and Horizontal Flipping as they were the most sensible
Pneumonia patient. choices for this type of data. We also resized all the images
(both training and testing) to 300x300 at this stage, to maintain
Observation: We observe that when compared to a normal X- uniformity at the time of training and testing. We chose
ray, in the thresholded image - the lung boundaries are very 300x300 because it is the average dimension of all the images
clearly defined here. We observe that even though for most and hence there was minimal loss of information. We used
images, this approach works well to demarcate the lung ImageDataGenerator function from Keras library for
boundary, important features are lost in some transformations. augmentation and resizing.
More precisely the boundary of the right lung is lost after the
Example -
transformation. It was observed [empirically] that
approximately only 1-2 images out of 15 were affected by this
and so training a model with this could be worthwhile. After
training on the thresholded images, the model could only
achieve an accuracy of 87%. We believe that the model did not
perform well because, in some images on the dataset, the
thresholding transform caused a large portion of the image to
turn white, these regions included the lower end of the lungs
which are a pivotal source of information for detection of
Pneumonia.
Morphological Operators: Figure 11 – Augmented Chest Xray with (A) 30% zoom and (B)
Horizontal Flip

2.5 TRAINING
We have used Convolutional Neural Networks for training
the model. We used the concept of Transfer Learning [13], it
is a technique in which a model created for one task is used as
the basis for a model for a different task. Usually, it is a model
which is trained on millions of images and whose
classification accuracy was found to be very high for a large
number of objects. In our case, the top layer of the model is
Figure 10 – Chest X-ray images of (a) Normal and (b) Pneumonia Resnet50v2 which is trained on ImageNet, followed by groups
patient after applying several Morphological Operations of dense and dropout layers to modify the output according to
our requirement. We have used ‘Adam’ as the optimizer and
Observation: From the above outputs, we can observe that ‘binary cross-entropy’ as the loss function because we are
when using erosion and dilation techniques, the images give us classifying binary (two) classes. We have used the concept of
exactly what we need to detect Pneumonia i.e. the edges of the ReduceLROnPlateau which reduces the learning rate (which
lungs are captured [especially near the bottom right], and the was initialized as 0.0001) if the validation loss doesn’t go
ribcage is also suppressed. down for a given number of epochs (in our case 5). We have
Conclusion: also used the concept of Early Stopping to avoid the model to
overfit by stopping the training after it reaches 97.5% training
After rigorous testing, our final pipeline was as follows: We accuracy hence instead of just mugging up the images, it
converted all images to grayscale and then used histogram learns the features from them. Hence our training accuracy
equalization followed by erosion, dilation and erosion was 97.5%. We used a batch size of 64 for training and testing
morphological operations (Also known as Opening then because we had a relatively large number of images and hence
it was possible to learn accurately while maintaining good

978-1-7281-9537-7/21/$31.00 ©2021 IEEE 70

Authorized licensed use limited to: International Islamic University Malaysia. Downloaded on July 20,2022 at 08:25:36 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the International Conference on Artificial Intelligence and Smart Systems (ICAIS-2021)
IEEE Xplore Part Number: CFP21OAB-ART; ISBN: 978-1-7281-9537-7

training speed. We also shuffled the training data as shuffling The definitions with equations of the three metrics:
data helps in reducing variance and making sure that the
1. Accuracy- It is the fraction of predictions our model got
model remains general and overfit less.
right i.e., number of correct predictions out of all predictions.
Intermediate activations of what our model is learning in its
black box:

(1)
2. Recall- It is the proportion of actual positives that were
classified correctly from all predictions.

(2)
3. Precision- It is the proportion of actual positives out of all
predicted positives.

Figure 12 - 2nd Convolutional Layer of our model visualized (3)

The percentage of correctly identified positive cases to the
overall number of positive cases is determined by recall. In the
medical domain, a False positive (A patient without Pneumonia
is diagnosed with it) is not as harmful as a False Negative (The
patient has Pneumonia, but it goes undiagnosed). Hence, recall
for us is the most important metric among these.

4. RESULT
In this paper, we propose a novel methodology that is most
unique in the preprocessing pipeline used on the X-ray images.
Our proposed study achieved an Accuracy of 96%, a Precision
score of 96.20%, a Recall score of 97.4% and an F1 score of
96.8% on the test data consisting of 624 images. The test data
Figure 13 - 25th Convolutional Layer of our model visualized
predictions had more False Positives than False Negatives
2.6 TESTING which means less Pneumonia suffering patients were classified
We tested our model against 624 test images. We processed as healthy, which is ideal in medical diagnosis. These results
all the test images as per our preprocessing pipeline and then proved the efficacy of the proposed method and confirmed that
evaluated our model against these images. We achieved a our method can be used for assisting diagnosis of Pneumonia.
classification accuracy of 96%, a precision score of 96.20% Here is the confusion matrix of our model for test images:
and a recall score of 97.4%, validating the model's
classification strength.

3. PERFORMANCE METRICS
We measured the performance of our model using three
metrics – Accuracy, Recall and Precision.
Meaning of abbreviations used in the equations below:
True positive (TP): Prediction is positive and original label is
also positive (pneumonia).
True negative (TN): Prediction is negative and original label
is also negative (healthy).
False positive (FP): Prediction is positive but original label is
negative (healthy).
False negative (FN): Prediction is negative but original label is Figure 14 – The Confusion Matrix
positive (pneumonia).

978-1-7281-9537-7/21/$31.00 ©2021 IEEE 71

Authorized licensed use limited to: International Islamic University Malaysia. Downloaded on July 20,2022 at 08:25:36 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the International Conference on Artificial Intelligence and Smart Systems (ICAIS-2021)
IEEE Xplore Part Number: CFP21OAB-ART; ISBN: 978-1-7281-9537-7

As we can observe only 25 images out of 624 were classified VGG19. The test dataset used by them was quite small as
incorrectly and hence the model is very effective and efficient. compared to this paper and they did not use any augmentation
techniques. Vikash et al. [9] proposed an ensemble model
Here are the 25 incorrectly classified images:
using five different models pre-trained on ImageNet –
GoogLeNet, Inception V3, AlexNet, DenseNet121 and
ResNet18. They used different parameter values for each
model and achieved a test score of 96.40% on the ensemble
model. Rahman et al. [8] used four different transfer learning
models - AlexNet, ResNet18, DenseNet201, and SqueezeNet.
Three classifications schemes have been reported: normal vs.
pneumonia, bacterial vs. viral pneumonia, and normal,
bacterial, and viral pneumonia with 98%, 95%, and 93.3% test
accuracy respectively. Ayan et al. [7] used two transfer
learning models - Vgg16 and Xception and achieved an
accuracy of 87% and 82% respectively. Although Xception
was less accurate it was more successful in detecting
pneumonia cases. Saraiva et al. [6] compared two neural
networks- Multilayer Perceptron and Convolution Neural
Network. Using K-fold cross-validation they reported 92.16%
and 94.40% accuracy on test data respectively.
As compared to all the papers above, the dataset used in this
paper is bigger because of upsampling and hence our model is
comparatively less biased. In this paper, the results for various
metrics are not the highest but this paper stands out by
Figure 15 – Incorrectly classified images providing a novel Image Processing based pipeline consisting
of Histogram Equalization and Morphological Operations to
On further analysis, we found out that these 25 images were detect Pneumonia and shows the potential to which Image
either Left Lingular Pneumonia or Right lower lobe Pneumonia Processing can transform results.
and hence the accuracy may be improved by addition of more
data of these type of Pneumonia in future works.
5. CONCLUSION
Details of various existing methods for Pneumonia Detection Pneumonia causes a considerable amount of morbidity and
are as follows: mortality. It has a high mortality rate of around 25% [10].
Rajaraman et al. [4] used customized VGG-16 and reported an World Health Organization says pneumonia can be prevented
accuracy of 96.2% and 93.6% for bacterial and viral with early diagnosis and treatment [12] but there is a huge
pneumonia respectively. They applied novel visualization shortage of experienced radiologists all across the globe,
strategies to explain model predictions and hence helping to especially in South Asia and Sub-Saharan Africa who can
make better clinical decisions. Cohen et al. [3] used the diagnose Pneumonia from X-rays. It is when projects like ours
DenseNet-121 architecture, which Rajpurkar et al. [21] become useful for helping and assisting doctors for more
demonstrated to function well on chest X-rays. He achieved an accurate diagnosis of the disease. In this paper, we proposed a
AUC score of 98.4% but this metric was not used for this technique for pneumonia detection from chest X-rays using
paper. Unfortunately, no other metrics were reported in the deep learning and image processing. Many experiments were
paper. Stephen et al. [2] used a custom Convolutional Neural performed with various image processing techniques, and a
Network model which they built from scratch and hence very effective pipeline was devised at the end which obtained
displayed high accuracy without using transfer learning. They high scores on various metrics and proved the robustness of the
reported an accuracy of 93.73% and did not publish any other model. Augmentation techniques were employed to avoid our
metric. M.Togacar et al. [5] used three deep learning models- model from overfitting. Transfer Learning was used to make up
AlexNet, VGG-16, VGG-19 and then used feature selection the base model because it saves a lot of training time, gives
method (minimum redundancy maximum relevance algorithm) better performance (in most cases), and doesn’t need a lot of
to reduce features from 1000 to 100 for each model and data.
achieved an accuracy of 96.84% by using 300 deep features. In conclusion, Image processing is a very effective technique
Rahib H. Abiyey et al. [19] did a comparative analysis of which can be used to drastically improve the performance of
custom-designed Backpropagation neural network (BPNN), deep learning models in most of the cases and hence it is the
Convolutional Neural Networks and Competitive Neural pipeline we used for Image processing which provides novelty
Networks (CpNN) and achieved an accuracy of 80.04%, to our method. Without image processing, it seemed that our
92.4%, 89.57% on test data respectively. They also compared model was learning quite less from the images as compared to
the performance of their customs models with VGG16 and when Image processing was applied. It was used to suppress

978-1-7281-9537-7/21/$31.00 ©2021 IEEE 72

Authorized licensed use limited to: International Islamic University Malaysia. Downloaded on July 20,2022 at 08:25:36 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the International Conference on Artificial Intelligence and Smart Systems (ICAIS-2021)
IEEE Xplore Part Number: CFP21OAB-ART; ISBN: 978-1-7281-9537-7

the ribcage and bring essential regions of the image into focus, [13] Niklas Donges (2019, June 16). What is transfer learninig? Exploring
the popular deep learning approach. Builltin. https://builtin.com/data-
thus making it possible to train a highly accurate convolutional science/transfer-learning
neural network model to detect the presence of Pneumonia.
[14] Our World in Data . “Number of people dying from pneumonia by age”.
It would be fascinating to see approaches consisting of various “Pneumonia”. https://ourworldindata.org/pneumonia#burden-of-
other types of image preprocessing pipelines and approaches pneumonia (Accessed: Oct. 2 2020)
involving weighted classifiers consisting of more than one [15] Mohammad Ali Kadampur, Sulaiman Al Riyaee. Skin cancer detection:
model being trained on these pipelines in the future. Applying a deep learning based model driven architecture in the cloud
for classifying dermal cell images. Current Issues in Informatics in
Medicine Unlocked, volume 18, 2020, 100282. Available:
REFERENCES https://doi.org/10.1016/j.imu.2019.100282.
[1] GLOWM. ”X-ray Atlas: Chest X-ray”. “The Global Library of [16] Bhandari, A., Koppen, J. & Agzarian, M. Convolutional neural networks
Women’sMedicine”. for brain tumour segmentation. Current Issues in Insights Imaging 11, 77
https://www.glowm.com/atlas_page/atlasid/chestXray.html (2020). Available: https://doi.org/10.1186/s13244-020-00869-4
[17] R., Dr. (2020). Deep Net Model for Detection of Covid-19 using
[2] Stephen O., Sain M., Maduh U.J., Jeong D.U. An efficient deep learning Radiographs based on ROC Analysis. Journal of Innovative Image
approach to pneumonia classification in healthcare. J. Healthc. Eng. Processing.2,no.03(2020):135-140.
2019;2019:4180949. doi: 10.1155/2019/4180949. [PMC free article]
[PubMed] [CrossRef] [Google Scholar] [18] Kallianos K., Mongan J., Antani S., Henry T., Taylor A., Abuya J.,
Kohli M. How far have we come? Artificial intelligence for chest
[3] Cohen J.P., Bertin P., Frappier V. Chester: A Web Delivered Locally radiograph interpretation. Clin. Radiol. 2019;74:338–345. doi:
Computed Chest X-Ray Disease Prediction System. arXiv. 10.1016/j.crad.2018.12.015. [PubMed] [CrossRef] [Google Scholar]
20191901.11210 [Google Scholar]
[19] Abiyev R.H., Ma’aitah M.K.S. Deep convolutional neural networks for
[4] Rajaraman S., Candemir S., Kim I., Thoma G., Antani S. Visualization chest diseases detection. J. Healthc. Eng. 2018;2018:4168538. doi:
and interpretation of convolutional neural network predictions in 10.1155/2018/4168538. [PMCfreearticle] [PubMed] [CrossRef] [Google
detecting pneumonia in pediatric chest radiographs. Appl. Sci. Scholar]
2018;8:1715. doi: 10.3390/app8101715. [PMC free article] [PubMed]
[CrossRef] [Google Scholar] [20] Pneumonia in the Western Pacific; Available Online:
https://www.who.int/westernpacific/health-topics/pneumonia
[5] Toğaçar M., Ergen B., Cömert Z. A deep feature learning model for
pneumonia detection applying a combination of mRMR feature selection [21] Pranav Rajpurkar, Jeremy Irvin, Kaylie Zhu, Brandon Yang, Hershel
and machine learning models. IRBM. 2019 doi: Mehta, Tony Duan, Daisy Ding, Aarti Bagul, Curtis Langlotz, Katie
10.1016/j.irbm.2019.10.006. [CrossRef] [Google Scholar] Shpanskaya, Matthew P. Lungren, and Andrew Y. Ng. CheXNet:
Radiologist-Level Pneumonia Detection on Chest X-Rays with Deep
[6] Saraiva A., Santos D., Costa N.J.C., Sousa J.V.M., Ferreira N.F., Learning. arxiv, 2017. Available Online:
Valente A., Soares S. Models of Learning to Classify X-ray Images for
the Detection of Pneumonia using Neural Networks; 2019 doi: https://arxiv.org/abs/1711.05225
10.5220/0007346600760083. Available online:
https://www.semanticscholar.org/paper/Models-of-Learning-to-Classify-
X-ray-Images-for-the-Saraiva-
Santos/0b8f202505b3d49c42fd45d86eca5dbd0b76fded?p2df.
[CrossRef]
[7] Ayan E., Ünver H.M. Diagnosis of Pneumonia from Chest X-Ray
Images Using Deep Learning; Proceedings of the 2019 Scientific
Meeting on Electrical-Electronics & Biomedical Engineering and
Computer Science (EBBT); Istanbul, Turkey. 2–26 April 2019; pp. 1–5.
[Google Scholar]
[8] Rahman T., Chowdhury M.E., Khandakar A., Islam K.R., Islam K.F.,
Mahbub Z.B., Kadir M.A., Kashem S. Transfer Learning with Deep
Convolutional Neural Network (CNN) for Pneumonia Detection using
Chest X-ray. Appl. Sci. 2020;10:3233. doi: 10.3390/app10093233.
[CrossRef] [Google Scholar]
[9] Chouhan V., Singh S.K., Khamparia A., Gupta D., Tiwari P., Moreira
C., Damaševičius R., de Albuquerque V.H.C. A Novel Transfer
Learning Based Approach for Pneumonia Detection in Chest X-ray
Images. Appl. Sci. 2020;10:559. doi: 10.3390/app10020559. [CrossRef]
[Google Scholar]
[10] Healthcare, University of Utah. Pneumonia Makes List for Top 10
Causes of Death;2016 Available online: https://healthcare.utah.edu/the-
scope/shows.php?shows=0_riw4wti7
[11] WHO Pneumonia is the Leading Cause of Death in Children. ;2011
Available online:
https://www.who.int/maternal_child_adolescent/news_events/news/2011
/pneumonia/en
[12] World Health Organization . Standardization of Interpretation of Chest
Radiographs for the Diagnosis of Pneumonia in Children. World Health
Organization; Geneva, Switzerland: 2001. Technical Report. [Google
Scholar]

978-1-7281-9537-7/21/$31.00 ©2021 IEEE 73

Authorized licensed use limited to: International Islamic University Malaysia. Downloaded on July 20,2022 at 08:25:36 UTC from IEEE Xplore. Restrictions apply.