Pneumonia Detection Using Convolutional Neural Networks (CNNs)

Abstract

Pneumonia, an interstitial lung disease, is the leading cause of death in children under the age of five. It accounted for approximately 16% of the deaths of children under the age of five, killing around 880,000 children in 2016 according to a study conducted by UNICEF. Affected children were mostly less than two years old. Timely detection of pneumonia in children can help to fast-track the process of recovery. This paper presents convolutional neural network models to accurately detect pneumonic lungs from chest X-rays, which can be utilized in the real world by medical practitioners to treat pneumonia. Experimentation was conducted on Chest X-Ray Images (Pneumonia) dataset available on Kaggle. The first, second, third and fourth model consists of one, two, three and four convolutional layers, respectively.
The first model achieves an accuracy of 89.74%, the second one reaches an accuracy of 85.26%, the third model achieves an accuracy of 92.31%, and lastly, the fourth model achieves an accuracy of 91.67%. Dropout regularization is employed in the second, third and fourth models tominimize overfitting in the fully connected layers.
Furthermore, recall and F1 scores are calculated from the confusion matrix of each model for better evaluation.

Full text

Pneumonia Detection Using
Convolutional Neural Networks (CNNs)
V. Sirish Kaushik, Anand Nayyar, Gaurav Kataria and Rachna Jain
Abstract Pneumonia, an interstitial lung disease, is the leading cause of death in
children under the age of five. It accounted for approximately 16% of the deaths of
children under the age of five, killing around 880,000 children in 2016 according to
a study conducted by UNICEF. Affected children were mostly less than two years
old. Timely detection of pneumonia in children can help to fast-track the process
of recovery. This paper presents convolutional neural network models to accurately
detect pneumonic lungs from chest X-rays, which can be utilized in the real world by
medical practitioners to treat pneumonia. Experimentation was conducted on Chest
X-Ray Images (Pneumonia) dataset available on Kaggle. The first, second, third and
fourth model consists of one, two, three and four convolutional layers, respectively.
The first model achieves an accuracy of 89.74%, the second one reaches an accuracy
of 85.26%, the third model achieves an accuracy of 92.31%, and lastly, the fourth
model achieves an accuracy of 91.67%. Dropout regularization is employed in the
second, third and fourth models to minimize overfitting in the fully connected layers.
Furthermore, recall and F1 scores are calculated from the confusion matrix of each
model for better evaluation.
Keywords Convolutional neural networks (CNNs) ·Pneumonia detection ·
ReLU ·Max-pooling ·Forward and backward propagation
V. Sirish Kaushik (B)·G. Kataria ·R. Jain
Bharati Vidyapeeth’s College of Engineering, New Delhi, Delhi, India
e-mail: shirishkaushik@gmail.com
G. Kataria
e-mail: gaurav.kataria2291999@gmail.com
R. Jain
e-mail: rachna.jain@bharatividyapeeth.edu
A. Nayyar
Graduate School, Duy Tan University, Da Nang, Vietnam
e-mail: anandnayyar@duytan.edu.vn
© Springer Nature Singapore Pte Ltd. 2020
P. K. Singh et al. (eds.), Proceedings of First International Conference on Computing,
Communications, and Cyber-Security (IC4S 2019), Lecture Notes in Networks
and Systems 121, https://doi.org/10.1007/978-981- 15-3369- 3_36
471

472 V. Sirish Kaushik et al.
1 Introduction
One of the major factors associated with pneumonia in children is indoor air pollution.
Apart from this, under-nutrition, lack of safe water, sanitation and basic health
facilities are also major factors. Pneumonia is an interstitial lung disease caused
by bacteria, fungi or viruses. It accounted for approximately 16% of the 5.6 million
under-five deaths, killing around 880,000 children in 2016 [1]. Affected victims were
mostly less than two years old. Timely detection of pneumonia can help to prevent
the deaths of children. This paper presents convolutional neural network models to
accurately detect pneumonic lungs from chest X-rays, which can be utilized in the
real world by medical practitioners to treat pneumonia [2]. These models have been
trained to classify chest X-ray images into normal and pneumonia in a few seconds,
hence serving the purpose of early detection of pneumonia. Although transfer
learning models based on convolutional neural networks like AlexNet, ResNet50,
InceptionV3, VGG16 and VGG19 are some of the most successful ImageNet dataset
models with pre-trained weights, they were not trained on this dataset as the size of
dataset taken for our research is not as extensive compared to ones which generally
employ transfer learning [3]. Four classification models were built using CNN to
detect pneumonia from chest X-ray images to help control this deadly infection in
children and other age groups. Accuracy of the model is directly correlated with the
size of the dataset, that is, the use of large datasets helps improve the accuracy of the
model, but there is no direct correlation between the number of convolutional layers
and the accuracy of the model.
To obtain the best results, a certain number of combinations of convolution layers,
dense layers, dropouts and learning rates have to be trained by evaluating the models
after each execution. Initially, simple models with one convolution layer were trained
on the dataset, and thereafter, the complexities were increased to get the model that
not only achieved desired accuracies but also outperformed other models in terms
of recall and F1 scores. The objective of the paper is to develop CNN models from
scratch which can classify and thus detect pneumonic patients from their chest X-
rays with high validation accuracy, recall and F1 scores. Recall is often favored in
medical imaging cases over other performance evaluating parameters, as it gives
a measure of false negatives in the results. The number of false negatives in the
result is very crucial in determining the real-world performance of models [4]. If a
model achieves high accuracy but low recall values, it is termed as underperforming,
inefficacious and even unsafe as higher false-negative values imply higher number of
instances where the model is predicting a patient as normal, but in reality, the person
is diseased. Hence, it would risk the patient’s life. To prevent this, the focus would
be only models with great recall values, decent accuracies and F1 scores [5].
The paper is organized into 5 sections: Sect. 1introduces the subject of this
research paper, addresses its importance and relevance, the purpose and motive to
undertake this research work and the objective of the paper. Section 2explores the
work related to this field that has been accomplished till now. Section 3explains
the methodology of the paper, explaining the architecture of the models, flowchart

Pneumonia Detection Using Convolutional … 473
and the dataset used to train and test the four models. Section 4presents the results
achieved by the various CNN models and compares the performance of each model
using accuracy and loss graphs and confusion matrices. Section 5provides a brief
conclusion to the paper and delivers the best-suited model. Furthermore, the future
scope of this research work has also been discussed. All the references which are
cited in the paper have been listed in the end.
2 Related Work
Many researchers have tackled the problem of classifying images with high accuracy.
Here are some citations related to our paper:
Rubin et al. [6] developed a CNN model to detect common thorax disease
from frontal and lateral chest X-ray images. MIMIC-CXR dataset was used to
perform large-scale automated recognition of these images. The dataset was split
into training, testing and validation sets as 70%, 20% and 10%, respectively. Data
augmentation and pixel normalization were used to improve overall performance.
Their DualNet CNN model achieved an average AUC of 0.72 and 0.688 for PA
and AP, respectively. A deep convolutional neural network to classify pulmonary
tuberculosis was developed by Lakhani et al. [7]. Transfer learning models such
as AlexNet and GoogleNet were also used to classify chest X-ray images. The
dataset was split into training, testing and validation sets as 68%, 14.9% and 17.1%,
respectively. Data augmentation and pre-processing techniques were employed to
get the best performing model achieving an AUC of 0.99. Precision and recall of the
model were 100 and 97.3%. An AG-CNN model was developed by Guan et al. [8]
to recognize thorax disease. ChestX-ray14 dataset was used to detect thorax disease
from chest X-ray images. Global and local branch attention-guided CNN was used
for classification purposes. Their model was better than other models mentioned
in their research paper, achieving an AUC of 0.868. A deep convolutional neural
network model was developed by Rajpurkar et al. [9] to classify chest X-ray images
into pneumonia and other 14 diseases. ChestX-ray14 dataset was used for training
the model. They compared their ChXNet model (121 layered model) with practicing
academic radiologists. Their ChXNet model achieved an F1 score (95% CI) of 0.435
outperforming radiologists which achieved an F1 score (95% CI) of 0.387.
A deep convolutional neural network model having five convolutional layers
some followed by max-pooling layers, having three fully connected layers was
trained by Krizhevsky et al. [10]. This network contained 60 million different
parameters. By employing dropout, this model achieved a top-five error percent
of 17%. Simonyan et al. [11] developed a highly accurate model employing multiple
small kernel-sized filters to achieve top-five test accuracy 92.7%. This model was
trained on the ImageNet dataset and submitted to the ILSVRC 2014 competition. A
convolution neural network for classification and segmentation of brain tumor MRIs
was developed by Xu et al. [12]. Multiple techniques such as data augmentation,
feature selection and pooling techniques were employed in this model. The validation

474 V. Sirish Kaushik et al.
accuracy for classification achieved by this model is 97.5%, and validation accuracy
of segmentation is 84%, 256 ×256 pixels sized frontal chest radiographs which were
fed to a deep convolution neural network to detect abnormalities. A convolutional
neural network with five convolution layers employing leaky ReLU, average pooling
and three fully connected layers was developed by Anthimopoulos et al. [13] to detect
interstitial lung disease patterns in a dataset containing 14,696 images belonging to
seven different classes. This model achieved a classification accuracy of 85.5%. He
et al. [14] developed a residual neural network (RNN) to classify images present in the
ImageNet dataset. RNN introduced the concept of shortcut connections to tackle the
problem of vanishing gradients. This model which was submitted to ILSVRC 2015
attained state-of-the-art classification accuracy. A transfer learning model, extension
of AlexNet using data augmentation techniques, was developed by Glozman et al.
[15]. This model was trained on ADNI database. Two neural network models were
presented by Hemanth et al. [16] which are MCPN and MKNN. These models
classified MRIs with high accuracies and tackled high convergence time period for
Artificial Neural Networks.
3 Methodology
CNN models have been created from scratch and trained on Chest X-Ray Images
(Pneumonia) dataset on Kaggle. Keras neural network library with TensorFlow
backend has been used to implement the models. Dataset consists of 5216 training
images, 624 testing images and 16 validation images. Data augmentation has been
applied to achieve better results from the dataset. The four models have been trained
on the training dataset, each with different number of convolutional layers. Each
model was trained for 20 epochs, with training and testing batch sizes of 32 and 1,
respectively. The following sub-headings further explain the above stages in depth.
3.1 CNN Architecture
CNN models are feed-forward networks with convolutional layers, pooling layers,
flattening layers and fully connected layers employing suitable activation functions.
Convolutional layer. It is the building block of the CNNs. Convolution operation
is done in mathematics to merge two functions [17]. In the CNN models, the input
image is first converted into matrix form. Convolution filter is applied to the input
matrix which slides over it, performing element-wise multiplication and storing the
sum. This creates a feature map. 3 ×3 filter is generally employed to create 2D
feature maps when images are black and white. Convolutions are performed in 3D
when the input image is represented as a 3D matrix where the RGB color represents
the third dimension. Several feature detectors are operated with the input matrix to
generate a layer of feature maps which thus forms the convolutional layer.

Pneumonia Detection Using Convolutional … 475
Activation functions. All four models presented in this paper use two different
activation functions, namely ReLU activation function and softmax activation
function. The ReLU activation function stands for rectified linear function [18].
It is a nonlinear function that outputs zero when the input is negative and outputs one
when the input is positive. The ReLU function is given by the following formula:
This type of activation function is broadly used in CNNs as it deals with the
problem of vanishing gradients and is useful for increasing the nonlinearity of layers.
ReLU activation function has many variants such as Noisy ReLUs, Leaky ReLUs
and Parametric ReLUs. Advantages of ReLU over other activation functions are
computational simplicity and representational sparsity. Softmax activation function
is used in all four models presented in this paper. This broadly used activation
function is employed in the last dense layer of all the four models [19]. This activation
function normalizes inputs into a probability distribution. Categorical cross-entropy
cost function is mostly used with this type of activation function.
Pooling layer. Convolutional layers are followed by pooling layers. The type of
pooling layer used in all four models is max-pooling layers. The max-pooling layer
having a dimension of 2 ×2 selects the maximum pixel intensity values from the
window of the image currently covered by the kernel. Max-pooling is used to down
sample images, hence reducing the dimensionality and complexity of the image [20].
Two other types of pooling layers can also be used which are general pooling and
overlapping pooling. The models presented in this paper use max-pooling technique
as it helps recognize salient features in the image.
Flattening layer and fully connected layers. After the input image passes
through the convolutional layer and the pooling layer, it is fed into the flattening
layer. This layer flattens out the input image into a column, further reducing its
computational complexity. This is then fed into the fully connected layer/dense layer.
The fully connected layer [21] has multiple layers, and every node in the first layer
is connected to every node in the second layer. Each layer in the fully connected
layer extracts features, and on this basis, the network makes a prediction [22,23].
This process is known as forward propagation. After forward propagation, a cost
function is calculated. It is a measure of performance of a neural network model.
The cost function used in all four models is categorical cross-entropy. After the cost
function is calculated, back propagation takes place. This process is repeated until
the network achieves optimum performance. Adam optimization algorithm has been
used in all four models.
Reducing overfitting. The first model exhibits substantial overfitting; hence,
dropout technique was employed in the later models [24]. Dropout technique helps to
reduce overfitting and tackles the problem of vanishing gradients. Dropout technique
encourages each neuron to form its own individual representation of the input data.
This technique on a random basis cuts connections between neurons in successive
layers during the training process [25]. Learning rate of models was also modified,
to reduce overfitting. Data augmentation technique can also be employed to reduce
overfitting.

476 V. Sirish Kaushik et al.
Algorithm of CNN classifiers. The algorithms used in the convolutional neural
network classifiers have been explained in Figs. 1and 2. Figure 3shows the flowchart
of the overall schema of research. The number of epochs for all the classifier models
presented in this paper was fixed at 20 after training and testing several CNN models
over the course of research. Classifier models trained for more number of epochs
have showed overfitting. Several optimizer functions were also trained and studied.
Adam optimizer function was finalized to be used for all classifiers after it gave the
best results. Initially, a simple classifier model with convolutional layer of image
size set to 64 * 64, 32 feature maps and employing ReLU activation function was
trained. Fully connected dense layer with 128 perceptrons was utilized. To improve
the result, the second classifier model was trained with one more convolutional
layer of 64 feature maps for better feature extraction. The number of perceptrons
in dense layer was also doubled to 256, so that better learning could be achieved.
The third model was trained for three convolutional layers with 128 feature maps in
third convolutional layer for more detailed feature extraction. Dense layer was kept
unchanged. Dropout layer was introduced at 0.3, and learning rate of optimizer was
Fig. 1 Algorithms of CNN classifier model 1 (left) and model 2 (right)

Pneumonia Detection Using Convolutional … 477
Fig. 2 Algorithms of CNN classifier model 3 (left) and model 4 (right)
lowered to 0.0001 to reduce the overfitting. The final fourth classifier model was
trained for four convolutional layers with 256 feature maps in fourth convolutional
layer. Dense layer, dropout layer and learning rate were kept same as third classifier
model. The results have been summarized in the subsequent section of this paper.
Dataset. Chest X-Ray Images (Pneumonia) dataset of 1.16 GB size has been
imported from Kaggle [26], with total of 5856 jpeg images split into Train, Test and
Val folders each divided into category Pneumonia and Normal. Chest X-ray images
(front and back) were selected from pediatric patients of one- to five-year olds from
Guangzhou Women and Children’s Medical Center, Guangzhou. Figure 4provides

478 V. Sirish Kaushik et al.
Fig. 3 Detailed schema of the experiment conducted

Pneumonia Detection Using Convolutional … 479
Fig. 4 Left image depicts normal lungs and right image depicts pneumonic lungs
the sample images from the dataset used during the research.
4 Experimental Results
To study the performance of each CNN classifier model, validation accuracy, recall
and F1 score were evaluated as the performance measures [27,28]. Accuracy and
loss graphs were also studied. The confusion matrix was also computed for each
model.
4.1 Comparison of Performance of Models
Figures 5and 6show the confusion matrices, accuracy graphs and loss graphs of
all CNN classifier models. Table 1and Figs. 5and 6show that classifier models
1 and 2 significantly underperformed compared to models 3 and 4. The accuracy
graphs and loss graphs show overfitting. Accuracy, recall and F1 scores are also low.
In addition to extra convolution layer, employing dropout and lowering the learning
rate of optimizer in model 3 improved the performance considerably. It achieved
the least overfitting along with highest accuracy and recall. Several attempts were
made to better the performance by adding more convolutional layers and changing
the parameters. Classifier model 4 with four convolutional layers showed good recall
value and F1 score albeit with lower accuracy and higher overfitting compared to
model 3. Thus, classifier model 3 performed the best among all CNN classifier
models. In the following equations, tp =true positive, tn =true negative, fp =false
positive and fn =false negative.
Accuracy =tp +tn
tp +tn +fp +fn (1)

480 V. Sirish Kaushik et al.
Fig. 5 Performance of classifier model 1 and model 2
Fig. 6 Performance of classifier model 3 and model 4
Precision =tp/(tp +fp)(2)

Pneumonia Detection Using Convolutional … 481
Tabl e 1 Performance comparison of different CNN models
Classifier model Validation accuracy
(%)
Validation loss (%) Recall (%) F1 score (%)
Model 1 (one
conv.layer)
89.74 27.31 96 92
Model 2 (two
conv.layers)
85.26 38.36 94 89
Model 3 (three
conv.layers)
92.31 25.23 98 94
Model 4 (4
conv.layers)
91.67 26.61 98 94
Recall =tp/(tp +fn)(3)
F1 Score =2(Precision ∗Recall)/(Precision +Recall)(4)
5 Conclusion
The validation accuracy, recall and F1 score of CNN classifier model 3 with three
convolutional layers are 92.31%, 98% and 94%, respectively, which are quite high
compared to other models that were trained. CNN classifier model 4 with four
convolutional layers also comes very close in performance with 91.67% validation
accuracy, 98% recall and 94% F1 score. Both of these models have the same recall and
F1 scores. The paper by Chakraborty [29] achieved the overall accuracy of 95.62%
and recall of 95% trained on the same dataset. The paper by Liang [30] achieved
recall of 96.7% on the same dataset. The models presented by us at best could
achieve 92.31% accuracy which is lower, but 98% recall has been achieved. High
recall values will ensure that the number of false-negative instances is lower, hence
lowers the risk to the patient’s life. Thus, it is concluded that CNN classifier model
3 and model 4 can, therefore, be effectively used by medical officers for diagnostic
purposes for early detection of pneumonia in children as well as adults. A large
number of X-ray images can be processed very quickly to provide highly precise
diagnostic results, thus helping healthcare systems provide efficient patient care
services and reduce mortality rates. These convolutional neural networks’ models
were successfully achieved by employing various methods of parameter tuning like
adding dropout, changing learning rates, changing the batch size, number of epochs,
adding more complex fully connected layers and changing various stochastic gradient
optimizers [31].

482 V. Sirish Kaushik et al.
In the future, it is hoped that transfer learning models would be trained on this
dataset that would outperform these CNN models. It is intended that larger datasets
will also be trained using the models presented in the paper. It is also expected that
neural network models based on GAN [32], generative adversarial networks, would
also be trained and compared with the existing models.
References
1. https://data.unicef.org/topic/child-health/pneumonia/. Accessed on 15 July 2019
2. Jaiswal, A.K., Tiwari, P., Kumar, S., Gupta, D., Khanna, A., Rodrigues, J.J.: Identifying
pneumonia in chest x-rays: a deep learning approach. Measurement 145, 511–518 (2019)
3. Kim, D.H., MacKinnon, T.: Artificial intelligence in fracture detection: transfer learning from
deep convolutional neural networks. Clin. Radiol. 73(5), 439–445 (2018)
4. Bernal, J., Kushibar, K., Asfaw, D.S., Valverde, S., Oliver, A., Martí, R., Lladó, X.: Deep
convolutional neural networks for brain image analysis on magnetic resonance imaging: a
review. Artif. Intell. Med. 95, 64–81 (2019)
5. Arthur, F., Hossein, K.R.: Deep learning in medical image analysis: a third eye for doctors. J.
Stomatology Oral Maxillofac. Surg.
6. Rubin, J., Sanghavi, D., Zhao, C., Lee, K., Qadir, A., Xu-Wilson, M.: Large Scale Automated
Reading of Frontal and Lateral Chest X-Rays Using Dual Convolutional Neural Networks
(2018). arXiv preprint arXiv:1804.07839
7. Lakhani, P., Sundaram, B.: Deep learning at chest radiography: automated classification of
pulmonary tuberculosis by using convolutional neural networks. Radiology 284(2), 574–582
(2017)
8. Guan, Q., Huang, Y., Zhong, Z., Zheng, Z., Zheng, L., Yang, Y.: Diagnose Like a Radiologist:
Attention Guided Convolutional Neural Network for Thorax Disease Classification (2018).
arXiv preprint arXiv:1801.09927
9. Rajpurkar, P., Irvin, J., Zhu, K., Yang, B., Mehta, H., Duan, T., Ding, D., Bagul, A., Langlotz,
C., Shpanskaya, K., Lungren, M.P.: Chexnet: Radiologist-Level Pneumonia Detection on Chest
X-rays with Deep Learning (2017). arXiv preprint arXiv:1711.05225
10. Krizhevsky, A., Sutskever, I., Hinton, G.E:. Imagenet classification with deep convolutional
neural networks. In: Advances in Neural Information Processing Systems, pp. 1097–1105
(2012)
11. Simonyan, K., Zisserman, A.: Very Deep Convolutional Networks for Large-Scale Image
Recognition (2014). arXiv preprint arXiv:1409.1556
12. Xu, Y., Jia, Z., Ai, Y., Zhang, F., Lai, M., Eric, I., Chang, C.: Deep convolutional activation
features for large scale brain tumor histopathology image classification and segmentation. In:
2015 international conference on acoustics, speech and signal processing (ICASSP), pp. 947–
951 (2015)
13. Anthimopoulos, M., Christodoulidis, S., Ebner, L., Christe, A., Mougiakakou, S.: Lung pattern
classification for interstitial lung diseases using a deep convolutional neural network. IEEE
Trans. Med. Imaging 35(5), 1207–1216 (2016)
14. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In:
Proceedings of the IEEE Conference on Computer Vision and PatternRecognition, pp. 770–778
(2016)
15. Glozman, T., Liba, O.: Hidden Cues: Deep Learning for Alzheimer’s Disease Classification
CS331B project final report (2016)
16. Hemanth, D.J., Vijila, C.K.S., Selvakumar, A.I., Anitha, J.: Performance improved iteration-
free artificial neural networks for abnormal magnetic resonance brain image classification.
Neurocomputing 130, 98–107 (2014)

Pneumonia Detection Using Convolutional … 483
17. Bi, X., Li, S., Xiao, B., Li, Y., Wang, G., Ma, X.: Computer aided Alzheimer’s disease diagnosis
by an unsupervised deep learning technology. Neurocomputing (2019)
18. Eckle, K., Schmidt-Hieber, J.: A comparison of deep networks with ReLU activation function
and linear spline-type methods. Neural Netw. 110, 232–242 (2019)
19. Ren, S., Jain, D.K., Guo, K., Xu, T., Chi, T.: Towards efficient medical lesion image super-
resolution based on deep residual networks. Sig. Process. Image Commun. 75, 1–10 (2019)
20. Zheng, Y., Iwana, B.K., Uchida, S.: Mining the displacement of max-pooling for text
recognition. Pattern Recogn. 93, 558–569 (2019)
21. Bhumika, P.S.S.S., Nayyar, P.A.: A review paper on algorithms used for text classification. Int.
J. Appl. Innov. Eng. Manage. 3(2), 90–99 (2013)
22. Kumar, A., Sangwan, S.R., Arora, A., Nayyar, A., Abdel-Basset, M.: Sarcasm detection using
soft attention-based bidirectional long short-term memory model with convolution network.
IEEE Access 7, 23319–23328 (2019)
23. Saeed, F., Paul, A., Karthigaikumar, P., Nayyar, A.: Convolutional neural network based early
fire detection. In: Multimedia Tools and Applications, pp. 1–17 (2019)
24. Kukkar, A., Mohana, R., Nayyar, A., Kim, J., Kang, B.G., Chilamkurti, N.: A novel deep-
learning-based bug severity classification technique using convolutional neural networks and
random forest with boosting. Sensors 19(13), 2964 (2019)
25. Khan, S.H., Hayat, M., Porikli, F.:Regularization of deep neural networks with spectral dropout.
Neural Netw. 110, 82–90 (2019)
26. https://www.kaggle.com/paultimothymooney/chest-xray-pneumonia. Accessed on 15 July
2019
27. ALzubi, J.A., Bharathikannan, B., Tanwar, S., Manikandan, R., Khanna, A., Thaventhiran, C.:
Boosted neural network ensemble classification for lung cancer disease diagnosis. Appl. Soft
Comput. 80, 579–591 (2019)
28. Vora, J., Tanwar, S., Polkowski, Z., Tyagi, S., Singh, P.K., Singh, Y.: Machine learning-based
software effort estimation: an analysis. In: 11th International Conference on Electronics,
computers and Artificial Intelligence (ECAI 2019), pp. 1–6, University of Pitesti, Pitesti,
Romania, 27–29 June 2019
29. Chakraborty, S., Aich, S., Sim, J.S., Kim, H.C.: Detection of pneumonia from chest x-rays using
a convolutionalneural network architecture. In: International Conference on Future Information
& Communication Engineering, vol. 11, no. 1, pp. 98–102 (2019)
30. Liang, G., Zheng, L.: A transfer learning method with deep residual network for pediatric
pneumonia diagnosis. In: Computer Methods and Programs in Biomedicine (2019)
31. Du, S. S., Zhai, X., Poczos, B., Singh, A.: Gradient Descent Provably Optimizes Over-
Parameterized Neural Networks (2018). arXiv preprint arXiv:1810.02054
32. Radford, A., Metz, L., Chintala, S.: Unsupervised Representation Learning with Deep
Convolutional Generative Adversarial Networks (2015). arXiv preprint arXiv:1511.06434