IAES International Journal of Artificial Intelligence (IJ-AI)

Vol. 13, No.4, December 2024, pp. 4169~4177

ISSN: 2252-8938, DOI: 10.11591/ijai.v13.i4.pp4169-4177  4169

Pneumonia detection on x-ray image using improved depthwise

separable convolutional neural networks

Islam Nur Alam1, Ghinaa Zain Nabiilah1, Erna Fransisca Angela Sihotang2, Bakti Amirul Jabar1
1
Department of Computer Science, School of Computer Science, Bina Nusantara University, Jakarta, Indonesia
2
Department of Statistics, School of Computer Science, Bina Nusantara University, Jakarta, Indonesia

Article Info ABSTRACT

Article history: A single neural network model cannot capture intricate and diverse features
due to its ability to learn only a finite set of patterns from the data.
Received Nov 13, 2023 Additionally, training and utilising a single model can be computationally
Revised Apr 5, 2024 demanding. Experts propose incorporating multiple neural network models to
Accepted Apr 17, 2024 address these constraints to extract complementary attributes. Previous
research has highlighted challenges network models face, including
difficulties in effectively capturing highly detailed spatial features,
Keywords: redundancy in network structure parameters, and restricted generalisation
capabilities. This study introduces an innovative neural network architecture
Chest x-ray that combines the Xception module with the inverse residue structure to tackle
Convolutional neural network these issues. Considering this, the paper presents a model for detecting
MobileNetV2 pneumonia in X-ray images employing an improved depthwise separable
Pneumonia convolutional network. This network architecture integrates the inverse
Xception residual structure from the MobileNetV2 model, using the rectified linear unit
(ReLU) non-linear activation function throughout the entire network. The
experimental results show an impressive recognition rate with a test accuracy
of 97.24% on the chest x-ray dataset. This method can extract more profound
and abstract image features while mitigating overfitting issues and enhancing
the network's generalisation capacity.
This is an open access article under the CC BY-SA license.

Corresponding Author:
Islam Nur Alam
Department of Computer Science, School of Computer Science, Bina Nusantara University
Jakarta 11480, Indonesia
Email: [email protected]

1. INTRODUCTION
Pneumonia, an infection of the lungs causing inflammation and fluid buildup, poses a significant
global health threat. This condition, characterized by shortness of breath and reduced oxygen intake, claims
the lives of countless children under five annually. In 2019 alone, worldwide figures show the death toll
reaching 740,180, with 314,455 of those young victims residing in Indonesia [1], [2]. These images can be
ambiguous, leading to misdiagnosis or conflicting interpretations, even for experienced radiologists.
Furthermore, manual analysis is time-consuming, potentially delaying critical treatment [3].
To address the challenges of interpreting chest X-rays, the implementation of a computerized system
emerges as a crucial solution to assist radiologists in identifying pneumonia. Machine learning, particularly
adept at image classification, exhibits promising potential in this domain. Chandra and Verma [4] exemplifies
this, where they evaluated the efficacy of various machine learning classifiers, including logistic regression,
multilayer perceptron, random forest, and sequential minimal optimization, in detecting pneumonia from
X-ray images [4]. Their findings provide compelling evidence of the capabilities of these models, laying the
groundwork for further exploration and optimization within this promising field.

Journal homepage: http://ijai.iaescore.com

4170  ISSN: 2252-8938

Beyond aiding diagnosis, the study employs feature extraction prior to classification, ultimately
achieving an impressive 95.53% accuracy with logistic regression, highlighting the potential of machine
learning approaches. Nevertheless, limitations arise when handling significant data volumes, as evidenced by
the study's use of a relatively small dataset (412 images) [5]. Deep learning offers a promising alternative for
overcoming these limitations, particularly in image classification domains like medical imaging. Convolutional
neural networks (CNNs) stand out as a popular and effective deep learning method [6], [7]. Their success in
detecting various medical issues like breast and lung cancer, brain tumors, and skin diseases underscores their
vast potential for pneumonia detection as well [8].
Leveraging transfer learning, a technique utilizing knowledge gained from one problem to solve
another, holds immense potential for addressing pneumonia detection challenges [9]. This approach capitalizes
on similarities between problems, accelerating learning in scenarios where acquiring data is difficult or costly,
like in medical contexts [10]. At its core, transfer learning bridges the gap between familiar and unfamiliar
information, fostering new insights. It entails transferring knowledge from a well-known "source domain" (e.g.,
general image classification) to the unfamiliar "target domain" (pneumonia detection in X-rays). The aim is to
explore effective methods for this knowledge transfer, allowing established models to apply previous learnings
to new information effectively. Within machine learning, transfer learning categorizes into four approaches
based on their methods: feature-based, model-based, relationship-based, and sample-based [11].
In the realm of deep neural networks, the training of models with a multilayer architecture using a
substantial amount of data is imperative for acquiring practical features and enhancing recognition accuracy.
While earlier studies have explored various aspects of deep learning architectures, critical examination reveals
certain gaps that need addressing. Specifically, the inadequacy of single network models in extracting intricate
and comprehensive features has been acknowledged. Additionally, the expansive structure of networks,
characterized by numerous parameters, poses challenges related to computational resources and efficiency.
This paper aims to address these gaps by introducing an enhanced Xception network model specifically tailored
for pneumonia detection in X-ray images.
The Xception model, conceived as an enhancement to Google's Inception-v3 architecture, employs
depthwise and pointwise convolutions to efficiently extract information from diverse channels and convolution
kernels. This innovative approach significantly reduces the number of parameters and associated computational
expenses. Furthermore, the modified Xception network structure incorporates the inverted residual design from
MobileNetV2, effectively mitigating issues related to gradient disappearance and explosion while enhancing
gradient propagation between layers. Recognizing the limitations inherent in neural network model training,
such as limited data availability and the risk of overfitting, our study employs data augmentation techniques to
augment the image dataset. However, it is important to acknowledge that the model outlined in this study is
not without its limitations, including suboptimal recognition rates for images afflicted by significant noise, low
resolution, and severe occlusion. Future research endeavors will continue to address these concerns,
contributing to the ongoing refinement and advancement of pneumonia detection using deep neural networks.

2. RELATED WORK
Building upon successful implementations of deep learning for pneumonia detection in chest X-ray
images, several studies have demonstrated promising advancements. Ayan and Ünver [12] achieved notable
results by leveraging transfer learning, fine-tuning, and data augmentation with adapted Xception and VGG16
models. Their work showcased the superior performance of the Xception model compared to VGG16. In 2021,
Zhang et al. [13] modified the VGG architecture, demonstrating its potential competitiveness among
established models like VGG-16, RES-50, Xception, DenseNet21, and MobileNet. These efforts highlight the
ongoing refinement of deep learning approaches for accurate pneumonia detection.
While prior studies achieved impressive results, our current study aims to explore ensemble stacking,
a technique that combines predictions from multiple models. Leveraging powerful architectures like Xception,
Resnet152V2, InceptionV3, VGG16, and VGG19, we have introduced a modification named multilevel
ensemble stacking. The primary goal is to further enhance the accuracy of diagnosing pneumonia through chest
X-rays.
Recent years have witnessed various methods for identifying pneumonia from chest X-ray images,
primarily using deep learning or deep CNN approaches. Rajpurkar et al. [14] employed a 121-layer CNN
model, achieving success in detecting pneumonia among 14 different chest-related diseases. Rahman et al. [15]
utilized transfer learning with four deep learning algorithms, where DenseNet201 outperformed others,
achieving a 98% accuracy rate. Varshni et al. [16] explored alternative machine learning methodologies like
support vector machine (SVM), naive Bayes, k-nearest neighbors (KNN), and random forest in conjunction
with deep CNN models. The study highlighted DenseNet-169 with SVM, yielding an area under the ROC
curve (AUC) of 0.8002 [16].

Int J Artif Intell, Vol. 13, No. 4, December 2024: 4169-4177

Int J Artif Intell ISSN: 2252-8938  4171

In the realm of image classification, the latest research endeavors focus on enhancing accuracy through
ensemble learning, combining top-performing models. Chouhan et al. [17] achieved an accuracy rate of 96.4%
by combining deep CNN models, while Mabrouk et al. [18] merged vision transformer, MobileNetV2, and
DenseNet169 for an optimal accuracy of 93.91%. The model we designed introduces improvements by
leveraging concepts from GoogLeNet, Xception, and ResNet. We combine the Xception model with the inverted
residual structure, recognized as an efficient and influential deep learning framework. Additionally, the use of
the global average pooling layer at the end of the model contributes to enhancing its accuracy.

3. METHOD
The study begins by collecting data in the form of "jpg" formatted X-ray images. The next step
involves data preprocessing using image augmentation techniques, which are detailed in the "dataset setup"
chapter. Next, we construct a deep neural network model known as depthwise separable CNNs. The strategy
we apply to this model aims to separate channel and spatial correlations, with the goal of saving network
parameters and improving model performance. Initially, the network employs a depthwise-pointwise
convolution structure, where depthwise convolution is performed first, followed by pointwise convolution in
the second step. Subsequently, we conduct experiments by performing several initial hyperparameter
adjustments, namely batch size of 32, 100 epochs, and using the Adam optimizer with a learning rate of 0.001.
Once the experiments are completed, we evaluate the model and compare it with previous research through a
benchmarking process.

3.1. Datasets setup

The present investigation utilised a publicly available dataset developed by Mooney and obtained via
the online platform Kaggle [18]. The dataset comprises roughly 5,197 chest X-ray images, which have been
classified into two distinct categories: pneumonia and normal. The chest X-ray pictures utilised in this
investigation were selected from a cohort of prior instances involving paediatric patients aged one to five who
had received chest X-ray examinations as part of their routine clinical evaluations at the Guangzhou women
and children's medical centre in Guangzhou. The images were acquired via the anterior-posterior view, the
conventional method employed for obtaining images of this nature [19].
To guarantee the reliability of the studied data, a meticulous quality control screening process was
applied to the chest X-ray images, filtering out low-quality or incomprehensible ones. Two expert physicians
then independently evaluated the remaining images to validate the accuracy of the AI system's training data.
Additionally, a third expert assessed the dataset to minimise potential grading errors. In order to enhance the
dataset, a data augmentation technique was also employed for all of the datasets [20]. The augmented chest
X-ray dataset includes 5,948 train set images and 1,487 test set images. This collection features chest X-ray
images from patients with pneumonia and those without the condition. An example image from the dataset is
depicted in Figures 1(a) and 1(b).

(a) (b)

Figure 1. An illustration of a dataset image is provided as sample data of (a) pneumonia; total data: 4273
images and (b) non pneumonia; total data: 3162 images

3.2. Construction of the neural network model

Equations to streamline the network's weight count, simplify the model's structure, and enhance
accuracy, this paper's network model lessens parameters through partial connectivity and weight sharing,
displaying hierarchical expression characteristics. It integrates the conceptual structure of depthwise separable
Pneumonia detection on x-ray image using improved depthwise separable … (Islam Nur Alam)
4172  ISSN: 2252-8938

convolution. In the Xception network module, the process begins with a 1×1 convolution applied to the input
image. Subsequently, a 3×3 convolution is employed on each channel after the convolution stage, leading to
the consolidation of outcomes. In contrast to the Inception-v3 [21] network model, this approach amplifies
model efficiency without inflating complexity. Moreover, a constructed network model applies a residual
connection mechanism to address performance degradation and gradient vanishing issues, enabling the training
of deeper networks while upholding optimal performance. The schematic representation of the architectural
arrangement of the neural network with reduced weight is illustrated in Figure 2. Image data enters from the
left, progresses through the middle layer to extract features, and culminates in classification results through the
softmax function.

Figure 2. Neural network architecture diagram

Convolution is a way to extract features from images using a CNN. It works by sliding a small filter,
called a kernel, over the image and multiplying each pixel value in the kernel with its corresponding pixel
value. The sum of the products is then assigned to a new image called the feature map. Convolution can be
used to enhance the critical features of an image and reduce noise. It can also be used to reduce an image's size,
making it easier for the CNN to process. There are two types of convolution padding: valid and same. Valid
padding does not add any padding to the image, so the feature map is smaller than the original image [22]. The
same padding adds padding to the image so that the feature map is the same size as the original image. The
convolution operation is computed by the (1), where 𝑋𝑗𝐿 is the 𝑗 -th feature map unit of the 𝐿 -th layer, 𝑋𝑖𝐿−1 is
the 𝑖 -th input of the 𝐿 − 1 -th layer, 𝜃𝑖𝑗 Represents the convolution kernel, 𝑏 is the bias unit, and 𝘨(𝑥) is the
activation function.

𝑋𝑗𝐿 = 𝘨(∑𝑖 𝑋𝑖𝐿−1 𝜃𝑖𝑗 + 𝑏) (1)

The network model uses max pooling after the depthwise separable convolution layer to compress
features and extract the most important ones, simplifying the network structure and reducing the risk of
overfitting. Max pooling works by selecting the maximum value from a small region of the input feature map
[23]. This work employs global average pooling as a substitution for the fully connected layer to mitigate
overfitting and decrease the parameter count. Global average pooling involves aggregating the input's spatial
information, enhancing the network's resilience to spatial alterations [24], [25].

3.3. Depthwise separable convolution

The construction follows a strategy separating channel and spatial correlations to save network
parameters and improve network model performance. Initially, the network employs a depthwise-pointwise
convolution structure, undertaking depthwise convolution first and then another in the second step. These dual

Int J Artif Intell, Vol. 13, No. 4, December 2024: 4169-4177

Int J Artif Intell ISSN: 2252-8938  4173

convolution operations serve distinct functions, extracting features from diverse channels with varying
convolution kernels. Maintaining the data integrity, the two convolution actions have no interceding nonlinearity
(ReLu), as presented in Figure 3. Initially, a 1×1 channel correlation convolution is executed, followed by a 3×3
convolution with the same output channels. This two-step feature extraction technique minimises the space and
time expenses in constructing and training the network while achieving image feature extraction comparable to
standard convolution. It significantly reduces computational load while upholding neural network accuracy.

Figure 3. Decoupling cross-channel and spatial correlation

3.4. Network architecture

This paper presents an improved technique for detecting pneumonia in X-ray images through a
depthwise separable convolutional network. The network's design integrates feature recognition inspired by
the structure of Xception [26]. They mainly emphasise the consecutive handling of features from three channels
after their fusion to improve the precision of detecting pneumonia. The network's structure and detailed
parameter information are depicted in Figure 2 and Table 1. Initially, it entails successive 2D convolution
procedures spanning two layers. The convolution kernel is 3×3, featuring 32 and 64 individual convolution
kernels with a stride of 1. Following this, the output experiences a series of stages involving residual
module 1, an inverted residual microstructure, residual module 1, residual module 2, and a subsequent sequence
of residual module 1, an inverted residual microstructure. The quantity of convolution kernels in these
structures varies: 128, 256, 256, 364, 364, and 512. The depthwise separable convolution employs a 3×3 kernel,
while MaxPooling uses a 3×3 kernel, and the 2D convolution layer uses a 1×1 kernel. Residual module 1
operates with a step size of 2, whereas the others use a step of 1. The output moves through a depthwise
separable convolution layer with a 3×3 kernel, 728 kernels, and a step size 1. Eventually, the output undergoes
a global average pooling layer and Softmax classifier. All 2D convolution and depthwise separable convolution
operations pass through batch normalization and ReLU activation layers to expedite the convergence of the
network structure and augment the extraction of non-linear features.

Table 1. Network structure and parameter table

Input Layer 𝑐 𝑠
224×224×3 Conv2D 32 1
224×224×32 Conv2D 64 1
112×112×64 Residual module 1 128 2
28×28×128 Inverted residual structure × 4 256 1
28×28×256 Residual module 1 256 2
14×14×256 Residual module 2 364 1
14×14×364 Residual module 1 364 2
7×7×364 Inverted residual structure × 2 512 1
7×7×512 SeparableConv2D 728 1
7×7×728 GlobalAveragePooling2D - -
1×1×728 Conv2D 2 -

Pneumonia detection on x-ray image using improved depthwise separable … (Islam Nur Alam)
4174  ISSN: 2252-8938

4. RESULTS AND DISCUSSION

This research centres on detecting pneumonia from static X-ray images. The investigation employed
a batch size of 32, an initial learning rate set at 0.001, and utilised the Adam optimiser to refine the training
process. Additionally, the network incorporated a width multiplier, α, set to 1. This integration of parameter α
aims to reduce the number of network parameters, a strategy observed to yield more favourable results than
reducing the model's depth.
Based on the findings in Figure 4, it can be seen that the network model is constructed by combining
the Xception model and the inverse residual structure correlated with separable CNN. The method proposed in
this study tends to have an accuracy proportion reaching 0.9724, which is much higher than the method used
by ensemble model: DenseNet 169, MobileNetV2, and vision transformer. Our study demonstrates that
network models with stronger robustness and higher generalization ability are not necessarily correlated with
poor performance on deep neural networks. The proposed method can leverage the improved Xception network
model without negatively impacting the single network model. Although models with more complex and rich
features, and very large network structures, have many parameters and consume more space resources, this
method is able to reduce the number of parameters and computational costs.

Figure 4. The training and testing accuracy and loss with chest X-ray images dataset

This study comprehensively explores the integration of the Xception model with an inverted residual
structure introduced into the network architecture. Integrating this model with the improved Xception network
structure provides a solution to the challenges of gradient vanishing and gradient explosion, while also
enhancing the gradient propagation capability between product layers. Nevertheless, some shortcomings were
found in the model built in this paper, such as low recognition rate for images with large noise, low resolution,
and serious occlusion. Therefore, further in-depth studies may be needed to confirm these findings, especially
in terms of image processing using specific methods before entering the feature extraction process by the
architecture. This is to further save the computation process and sharpen the image analysis.
Furthermore, derived from the findings of the confusion matrix chart depicted in Figure 5, a significant
element enhancing the model's performance to achieve precise outcomes comprises the true positive (TP) and
true negative (TN) values. TP refers to the volume of positive data accurately identified by the model. TN
represents the volume of harmful data correctly classified as unfavourable by the model. The more TP or TN
the model has, the higher the accuracy of the model. Both TP and TN values positively contribute to accuracy,
as they represent correct predictions by the model.

4.1. Comparative experiments of different methods

Table 2 provides a comparative analysis of accuracy for various models developed by different
researchers. Among the models mentioned, the paper's proposed model stands out with the highest accuracy
score of 97.24%. This implies that the model outperforms the alternatives presented in the table, including a
CNN model with oversampling, a proposed neural network with VGG16, a proposed ensemble CNN, and an
ensemble model using various architectures like DenseNet 169, MobileNetV2, and vision transformer. The
notable difference in accuracy signifies that the paper's model has substantially advanced in the field of study,
likely achieving better results. It attests to the effectiveness of the "improved depthwise separable convolutional
network" in achieving higher accuracy, a crucial metric in various machine learning and computer vision
applications.

Int J Artif Intell, Vol. 13, No. 4, December 2024: 4169-4177

Int J Artif Intell ISSN: 2252-8938  4175

Figure 5. Confusion matrix on chest x-ray images dataset

Table 2. Comparison of test accuracy with previous research

Reference Method Accuracy (%)
Liebenlito et al. [27] CNN Model with oversampling 95.00
Sharma and Guleria [28] Proposed NN with VGG16 92.15
Ayan et al. [29] Proposed ensemble CNN 95.83
Mabrouk et al. [18] Ensemble Model: DenseNet 169, MobileNetV2, and vision transformer 93.91
Paper Method Improved depthwise separable convolutional network 97.24

5. CONCLUSION
This research leverages the depthwise separable convolutional network architecture, which integrates
a residual network with an inverse residual structure. By addressing limitations inherent in conventional
algorithms for pneumonia detection in X-ray images, such as the inability to extract high-level depth features
and the weak generalization ability of a single deep network model, this paper introduces an innovative method
for pneumonia detection based on depth-resolved enhanced convolutional networks. The proposed approach
adeptly combines traditional feature extraction methodologies with artificial neural networks, enabling the
capture of more profound and abstract image features. This results in a reduction in the impact of lighting and
variations in lung images, among other factors. Through the integration of the Xception model with the inverse
residual structure, the network model successfully mitigates overfitting, addressing issues related to gradient
loss and overamplification. Empirical findings demonstrate the specific enhancement of pneumonia detection
accuracy by the proposed model, reinforcing the network's resilience and generalization ability. Suggestions
for future research include exploring the model's susceptibility to more complex variations and occlusions, as
well as more realistic clinical scenarios, such as variations in patient age and health conditions, and the
possibility of X-ray images having different levels of damage. Furthermore, the study can be extended to
consider larger datasets, incorporating data from diverse sources and origins to enhance the model's
generalizability beyond the training dataset. Additionally, to deepen the understanding of the characteristics of
X-ray images, considering more detailed analysis techniques, such as image segmentation for infected area
recognition, could be the next step in advancing this research.

ACKNOWLEDGEMENTS
The authors express their gratitude to the Department of Computer Science, School of Computer
Science Bina Nusantara University for their appreciated encouragement.

REFERENCES
[1] K. Kanwal, S. G. Khalid, M. Asif, F. Zafar, and A. G. Qurashi, “Diagnosis of Community-Acquired pneumonia in children using
photoplethysmography and machine learning-based classifier,” Biomed Signal Process Control, vol. 87, Jan. 2024, doi:
10.1016/j.bspc.2023.105367.
[2] A. M. Sobirovna, “Causes of pneumonia in children,” Science and Innovation: International Scientific Journal, vol. 3, 2024, doi:
10.5281/zenodo.10598855.
[3] O. Stephen, M. Sain, U. J. Maduh, and D.-U. Jeong, “An efficient deep learning approach to pneumonia classification in healthcare,”
Journal of Healthcare Engineering, vol. 2019, pp. 1–7, Mar. 2019, doi: 10.1155/2019/4180949.

Pneumonia detection on x-ray image using improved depthwise separable … (Islam Nur Alam)
4176  ISSN: 2252-8938

[4] T. B. Chandra and K. Verma, “Pneumonia detection on chest x-ray using machine learning paradigm,” 3rd International Conference
on Computer Vision and Image Processing: CVIP 2018, 2020, pp. 21–33, doi: 10.1007/978-981-32-9088-4_3.
[5] N. Sharma, V. Jain, and A. Mishra, “An analysis of convolutional neural networks for image classification,” Procedia Computer
Science, vol. 132, pp. 377–384, 2018, doi: 10.1016/j.procs.2018.05.198.
[6] P. Chagas et al., “Evaluation of convolutional neural network architectures for chart image classification,” in 2018 International
Joint Conference on Neural Networks (IJCNN), IEEE, Jul. 2018, pp. 1–8, doi: 10.1109/IJCNN.2018.8489315.
[7] Z. Wang et al., “Breast cancer detection using extreme learning machine based on feature fusion with CNN deep features,” IEEE
Access, vol. 7, pp. 105146–105158, 2019, doi: 10.1109/ACCESS.2019.2892795.
[8] W. Alakwaa, M. Nassef, and A. Badr, “Lung cancer detection and classification with 3D convolutional neural network (3D-CNN),”
International Journal of Advanced Computer Science and Applications, vol. 8, no. 8, 2017, doi: 10.14569/IJACSA.2017.080853.
[9] I. N. Alam, I. H. Kartowisastro, and P. Wicaksono, “Transfer learning technique with efficientnet for facial expression recognition
system,” Revue d’Intelligence Artificielle, vol. 36, no. 4, pp. 543–552, Aug. 2022, doi: 10.18280/ria.360405.
[10] J. C. Hung, K.-C. Lin, and N.-X. Lai, “Recognizing learning emotion based on convolutional neural networks and transfer learning,”
Applied Soft Computing, vol. 84, Nov. 2019, doi: 10.1016/j.asoc.2019.105724.
[11] F. Zhuang et al., “A comprehensive survey on transfer learning,” Proceedings of the IEEE, vol. 109, no. 1, pp. 43-76, Jan. 2021,
doi: 10.1109/JPROC.2020.3004555.
[12] E. Ayan and H. M. Unver, “Diagnosis of pneumonia from chest x-ray images using deep learning,” in 2019 Scientific Meeting on
Electrical-Electronics & Biomedical Engineering and Computer Science (EBBT), IEEE, Apr. 2019, pp. 1–5, doi:
10.1109/EBBT.2019.8741582.
[13] D. Zhang, F. Ren, Y. Li, L. Na, and Y. Ma, “Pneumonia detection from chest x-ray images based on convolutional neural network,”
Electronics, vol. 10, no. 13, Jul. 2021, doi: 10.3390/electronics10131512.
[14] P. Rajpurkar et al., “CheXNet: Radiologist-level pneumonia detection on chest x-rays with deep learning,” arXiv-Computer Science,
pp. 1-7, Nov. 2017, doi: 10.48550/arXiv.1711.05225.
[15] T. Rahman et al., “Transfer learning with deep convolutional neural network (CNN) for pneumonia detection using chest x-ray,”
Applied Sciences, vol. 10, no. 9, 2020, doi: 10.3390/app10093233.
[16] D. Varshni, K. Thakral, L. Agarwal, R. Nijhawan, and A. Mittal, “Pneumonia detection using CNN based feature extraction,” in
2019 IEEE International Conference on Electrical, Computer and Communication Technologies (ICECCT), IEEE, Feb. 2019, pp.
1–7, doi: 10.1109/ICECCT.2019.8869364.
[17] V. Chouhan et al., “A novel transfer learning based approach for pneumonia detection in chest x-ray images,” Applied Sciences,
vol. 10, no. 2, p. 559, Jan. 2020, doi: 10.3390/app10020559.
[18] A. Mabrouk, R. P. D. Redondo, A. Dahou, M. A. Elaziz, and M. Kayed, “Pneumonia detection on chest x-ray images using ensemble
of deep convolutional neural networks,” Applied Sciences, vol. 12, no. 13, Jun. 2022, doi: 10.3390/app12136448.
[19] D. S. Kermany et al., “Identifying medical diagnoses and treatable diseases by image-based deep learning,” Cell, vol. 172, no. 5,
pp. 1122-1131, Feb. 2018, doi: 10.1016/j.cell.2018.02.010.
[20] C. S. Won, “Multi-scale CNN for fine-grained image recognition,” IEEE Access, vol. 8, pp. 116663–116674, 2020, doi:
10.1109/ACCESS.2020.3005150.
[21] M. Lin, Q. Chen, and S. Yan, “Network in network,” arXiv-Computer Science, pp. 1-10, Dec. 2013, doi: 10.48550/arXiv.1312.4400.
[22] F. Chollet, “Xception: deep learning with depthwise separable convolutions,” 2017 IEEE Conference on Computer Vision and
Pattern Recognition (CVPR), Honolulu, HI, USA, 2017, pp. 1800-1807, doi: 10.1109/CVPR.2017.195.
[23] M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.-C. Chen, “MobileNetV2: inverted residuals and linear bottlenecks,” arXiv-
Computer Science, pp. 1-14, Jan. 2018, doi: 10.48550/arXiv.1801.04381.
[24] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” 2016 IEEE Conference on Computer Vision
and Pattern Recognition Deep, 2016, doi: 10.1109/CVPR.2016.90.
[25] C. Szegedy, S. Ioffe, V. Vanhoucke, and A. A. Alemi, “Inception-v4, inception-resnet and the impact of residual connections on
learning,” Thirty-First AAAI Conference on Artificial Intelligence, vol. 3, no. 1, 2017, doi: 10.1609/aaai.v31i1.11231.
[26] S. Li, H. Qu, X. Dong, B. Dang, H. Zang, and Y. Gong, “Leveraging deep learning and xception architecture for high-accuracy
MRI classification in alzheimer diagnosis,” arXiv-Electrical Engineering and Systems Science, pp. 1-9, Mar. 2024, doi:
10.48550/arXiv.2403.16212.
[27] M. Liebenlito, Y. Irene, and A. Hamid, “Classification of tuberculosis and pneumonia in human lung based on chest x-ray image
using convolutional neural network,” InPrime: Indonesian Journal of Pure and Applied Mathematics, vol. 2, no. 1, pp. 24–32, Mar.
2020, doi: 10.15408/inprime.v2i1.14545.
[28] S. Sharma and K. Guleria, “A deep learning based model for the detection of pneumonia from chest x-ray images using VGG-16
and neural networks,” Procedia Computer Science, vol. 218, pp. 357–366, 2023, doi: 10.1016/j.procs.2023.01.018.
[29] E. Ayan, B. Karabulut, and H. M. Ünver, “Diagnosis of pediatric pneumonia with ensemble of deep convolutional neural networks
in chest x-ray images,” Arabian Journal for Science and Engineering, vol. 47, no. 2, pp. 2123–2139, Feb. 2022, doi:
10.1007/s13369-021-06127-z.

BIOGRAPHIES OF AUTHORS

Islam Nur Alam is a lecturer at Bina Nusantara University (BINUS). He has two
years of experience as a data science researcher with a proven track record in building successful
algorithms and predictive models for image classification using convolutional neural networks.
He is highly proficient in clustering and classification, content-based filtering, data analysis and
visualisation. He also continues to hone individual skills in data science, mainly focusing on
natural language processing and machine translation tasks. He can be contacted at email:
[email protected].

Int J Artif Intell, Vol. 13, No. 4, December 2024: 4169-4177

Int J Artif Intell ISSN: 2252-8938  4177

Ghinaa Zain Nabiilah is a lecturer from Bina Nusantara University (BINUS). She
graduated from BINUS, Department of Computer Science in 2023. Since 2020, her research has
been related to natural language processing, especially in analysing human personality and
emotions. In addition, she is also active in research on the management and investigation of
toxic sentences, hoaxes, and hate speech for decision support. She can be contacted at email:
[email protected].

Erna Fransisca Angela Sihotang is a lecturer in Department of Statistics at School

of Computer Science in Bina Nusantara University. She finished her Bachelor's in Statistics at
Diponegoro University and then continued her Master's in Computer Science at BINUS
University. Her research interests are applied statistics, decision model, and statistical
computing. She can be contacted at email: [email protected].

Bakti Amirul Jabar is an accomplished academic and researcher with a Bachelor

of Science degree in Mathematics from Institut Teknologi Bandung and a Master's degree in
Computer Science from BINUS University. He currently serves as a lecturer at BINUS
University, where he imparts his knowledge in artificial intelligence. His research contributions
to AI have earned him a prominent place in the academic community, and he is actively involved
in advancing the understanding and application of AI. For inquiries or collaboration. He can be
contacted at email: [email protected].

Pneumonia detection on x-ray image using improved depthwise separable … (Islam Nur Alam)