2024 Second International Conference on Advance in Information Technology (ICAIT-2024)

A Convnet Model-Based Deep Learning

Approach for Pneumonia Recognition from
Chest X-ray Images
1st Venkateswarulu Naik. B
2nd Bhuvan Unhelkar 3rd Prasun Chakrabarti
Department Of Information Systems
Department Of Information Systems Department of Computer Science and
And Decision Sciences, Muma College
And Decision Sciences, Muma College Engineering
Of Business
Of Business Sir Padampat Singhania University,
University of South Florida,
University of South Florida, Udaipur Rajasthan,lndia
South Florida, USA
South Florida, USA [email protected]
[email protected]
[email protected]
4th G S Pradeep Ghantasala
5th Meghana Pyla
Department of Computer Science and 5th Pellakuri Vidyullatha
IEEE Member, Department of
Engineering, Department of Computer Science and
Computer Science and Engineering
Alliance College of Engineering and Engineering
Alliance University,
Design, Alliance University, Koneru Lakshmaiah Education
Bengaluru, India
Bengaluru, India Foundation,
[email protected].
[email protected] Vaddeswaram, Guntur, Andhra
in
Pradesh, India
[email protected]

979-8-3503-8386-7/24/$31.00 ©2024 IEEE

Abstract— Pneumonia diagnosis using chest X-ray (CXR) holds an immense potentiality for diagnosing pneumonia
images shows a decisive part in early detection and handling of the infections by augmentation processes and data-driven
disease. Within this research, proposes a convolutional neural approaches.
network (ConvNet) model for automated pneumonia diagnosis
based on CXR images. Our model leverages the deep learning
This literature search yielded a huge collection of
capabilities of ConvNets to analyze features in CXR images and
categorize them as pneumonia-positive or pneumonia-negative. We
exploration on ConvNet-based models for pneumonia
conducted experimentations on a comprehensive dataset of CXR infections to drawn the conclusions from CX-Ray pictures.
images, achieving auspicious outcomes in terms of efficiency and Various phases of preparation have confirmed the
accuracy. The proposed ConvNet model demonstrates prospective helpfulness of ConvNets in definitively recognizing
for an accurate and appropriate pneumonia diagnosis, contributing pneumonia-related irregularities, including penetrates and
to improved healthcare outcomes and resource optimization in solidifications, from chest X-beam pictures. Different
medical settings. ConvNet model is implemented using tensorflow models, like AlexNet, ResNet, VGG, and DenseNet, have
keras with preprocessing using Image data generator. The models been utilized and modified to advance execution for
are deployed like sequential, Conv2D with tuning parameters of
pneumonia discovery assignments [5]. Preprocessing
Relu, dropout, max pool 2d and adam optimizer. For training the
images with 150, 150 pixels, train sample as are 1000 and input concepts with augmenting techniques for transfer learning
shape is 3, epochs are 50, batch size is 20. Applications of the models that to move learning algorithms have additionally
categorical cross-entropy loss function for multi-class division with been investigated to improve model speculation and
Adam optimizer employed for parameter optimization. The Model implementation. Besides, studies have examined the mix of
trained for a stable amount of epochs through early based on clinical information and radiological discoveries with
validation loss toward prevent overfitting and learning rate, batch ConvNet-based models to advance the indicative exactness
size, and dropout rate optimized through hyperparameter tuning. and clinical administration. While ConvNet deep learning
Keywords— Chest X-ray Images, Pneumonia, Deep learning, models have shown promising outcomes in pneumonia
convolutional neural network, fully connected layer,
Hyperparameter tuning
findings from CX-Ray pictures, a few difficulties remain.
Restricted explained datasets, vague issues, and
I. Introduction interpretability of model expectations present huge obstacles
Pneumonia has a prevalent and possibly severe respiratory in real world arrangements [6]. Besides, there is a need for
condition pigeonholed through inflammation for lung tissue, predictable evaluation measurements and approval
often produced by viral, bacterial, or a fungal infections. conventions to guarantee the unwavering quality and
Timely and precise diagnosis of pneumonia is crucial for reproducibility of ConvNet-based approaches in clinical
timely intervention and operational patient management [1]. settings. Future exploration bearings might zero in on
Conventional diagnostic techniques, such as clinical tending to these difficulties through cooperative endeavors,
assessment and laboratory tests, are valuable but may lack enormous scope dataset curation, and progress of
the specificity and speed required for rapid diagnosis and interpretable profound learning models for pneumonia
the treatment initiation. conclusion [7]. Convolutional networks are more effective
Over the past few years, around has remained a increasing in modern tools for utilizing a chest X-beam to analyze
attentiveness in leveraging medical imaging procedures, pneumonia imaging, offering the possibility to upgrade
predominantly chest X-ray imaging, for diagnosing demonstrative exactness and work on persistent results.
pneumonia infections [2]. CX-Ray images provided the Notwithstanding challenges, persevering through
anatomical details with lot of information of chest cavity, movements in ConvNet-based models hold affirmation for
clinical enabling’s that to show the different types of working with early discovery and treatment of pneumonia
abnormalities indicates infections of pneumonia viz lung [8,20,21]. Preceded with research endeavors and
filter issues, consolidate and pleural effusions. Moreover, interdisciplinary coordinated efforts are fundamental for
interpreting with the CX-Ray images that to expertise the tackle the maximum capacity the utilization of profound
subject to scale up the variability in inter-observing which learning in helpful settings.
highlights the need of automated good accurate diagnostic This paper presents a deep learning method to classify
systems. Highly effective representations deep learning on pneumonia infections utilize to build a ConvNet model
Large and complex features of datasets can be trained using trained on radiograph images. We aim to probe the
Convolutional ConvNets models that to be build to identify usefulness of this deep learning procedure in a detecting
the objects and to segment the tasks to be classified. The pneumonia from imaging data, relating its presentation to
most recent significance on present research interests is to traditional diagnostic methods and assessing its potential
develop ConvNets on medical analysis particularly on clinical utility. Through a comprehensive analysis of
diagnosis of pneumonia from CX-Ray images [3].Deep ConvNet-based diagnosis, we seek to contribute to the
learning methods used the learning algorithms, that are progressing landscape of pneumonia diagnostics, informing
predominantly build the convolutional neural networks clinicians, researchers, and policymakers on the assimilation
models to demonstrate the significant capabilities to classify of AI technologies into routine clinical practice. Our study
the images and recognition the patterns. By leveraging aims to estimate the performance of a CNN model trained
dense layers of interconnected neurons, in CNN models that onto an enormous dataset of chest radiographs for
could autonomously acquire the variant features in input automated detection of pneumonia. By comparing the
dimensional image raw data that ai to determine complex diagnostic accurateness of the CNN model per conventional
patterns of infected pneumonia cases on chest radiographs methods, such as clinical assessment and radiological
with more percentage of accuracy and high rated efficiency interpretation by human experts, we seek to assess the
[4,19]. This transformative of deep learning technologies
potential clinical utility and impact of deep learning in learning for medical image analysis, this work is an
pneumonia diagnosis. invaluable tool for practitioners and researchers.
II. Literature study
This literature review aims to explore recent advancements Shin et al. (2016) studied [14] comprehensively on the
in pneumonia diagnosis using ConvNets with CXR images. utilization of deep learning with convolutional neural
We will examine key studies, methodologies, datasets, and networks (CNNs) for a computer-aided detection (CAD)
outcomes in this promptly changing field. Through existing tasks into the classification of medical images. This research
and investigation discoveries, the ConvNets as a reliable and explores in various aspects of different features of dataset
effective device for mechanized pneumonia determination, with CNN architectures, and transfer learning techniques
consequently working with early location, worked on quiet relevant to CAD applications. The authors investigate and
results, and asset enhancement in medical services settings. highlight the design and implementation of CNN models on
Various stages of preparation have found the utilization of medical image analysis, with architectural dimensions for
CNN models for pneumonia conclusion utilizing CX-Ray performance implications. Additionally, this research work
imageries. discusses the different features of medical image datasets,
Rajpurkar et al. (2017) introduced a new model like including size, shape, contrast, diversity, and annotation,
CheXNet, a deep learning method to train the complex picture quality that will impact on CNN model training and
dataset of CX-Ray images to detect infection rates of validation process. Furthermore, Shin et al. delve into
pneumonia. The model performed and achieved to transfer learning strategies, elucidating how pre-trained
comparing other models and demonstrated that automatical CNN models might be revised to precise CAD tasks through
identification of pneumonia infections [9]. Similarly, Wang fine-tuning and feature extraction. Through a thorough
et al. (2017) proposed [10] a framework based deep learning examination of CNN-based CAD methodologies, this paper
approach named DenseNet which is similar to VGG model provides insightful information about the difficulties and
for detection of pneumonia. They developed model with cutting-edge methods of using deep learning for medical
more specificity and sensitivity ratios and performed well imagery analysis.
compare to traditional AI-ML approaches than the Smith, M etal 2020, shows [15] the absenteeism of
effectiveness of deep learning methods on medical diagnosis physicians and diagnosis provision schemes in India
of pneumonia infections. Moreover, Zhang et al. (2018) obligates negatively impacted of health organization.
developed [11] a multi-scale CNN model with high dense Moreover, hospitals struggle to find radiologists,
layers for identification of pneumonia that incorporate with particularly in rural regions, and doctors are overburdened
all dimensions that are extracted from CX-Ray images. with patients. As a consequence, nearly entirely cases were
Their developed model shows the improvement in accuracy cured by one doctor, which outcomes in a high rate of
to differentiate among pneumonia infections and normal incorrect diagnoses. Systems for computer-assisted
image of CX-Ray images showcasing the more significance diagnosis are being created to deal with this issue. The
for extraction of features to represent the NN-based current study compared the approaches and outcomes of
approaches. several neural network-based pneumonia detection
Liu et al. (2019) analyzes [12] the relationship between's procedures. Papers consuming the identical data set Chest
radiological and pathological dimensions in diagnosing X-ray14 are the only ones studied for the finest assessments.
infections of pneumonia. In Thoracic Imaging, publication Johnson, R., et al. 2019, [16] Pneumonia remains a
this article analyzes into the specifics of pneumonia substantial global cause of morbidity and death, requiring
analysis, especially focusing on the connection between precise and effective diagnosis techniques for prompt
radiological imaging and neurotic assessment. Through a far intervention. This systematic review aims to compare
and wide investigation of existing survey, the qualities and various diagnostic methods for pneumonia, evaluating their
impediments of current demonstrative methodologies, accuracy, feasibility, and practicality in clinical settings.
featuring the significance of incorporating radiological and The literature survey highpoints the significant progress
neurotic information for precise finding and treatment made in employing Convolutional Neural Network models
arranging. The disclosures for audit guarantee to upgrading for a pneumonia diagnosis consuming Chest X-ray images.
our sympathetic of pneumonia finding and make ready for While CNNs offer promising results and automation
worked on clinical practices in the field of thoracic imaging. potential, addressing challenges such as dataset quality,
Park et al. (2018) provides [13] an overview that widespread model interpretability, and clinical integration is crucial for
on different applications of deep learning in medical advancing this field further.
synthesis the images. This research work surveys the present III. Model Building
situation landscape of deep learning techniques and but not Our Convolutional Neural Network (ConvNet) classical for
inadequate to magnetic resonance images alike X-ray and the pneumonia detection was designed toward automatically
tomography scans. The authors discussed the main concepts analyzes chest radiographs and classifies them into two
and classification models such as like as recurrent neural categories: pneumonia-positive and pneumonia-negative.
networks then convolutional neural networks with The architecture of our ConvNet model entails of multiple
segmentations and image detection. Moreover, this research layers, Using pooling layers to reduce dimensionality,
explores more usable challenges and opportunities in the convolutional layers which extract features, and fully
field of deep learning in clinical practices to emphasize the connected layers to categorize.
improved diagnosis with high percentage of accuracy and The input to the ConvNet model is a chest radiograph
more efficiency. Overall, when it comes to consuming deep image, represented as a two-dimensional array of pixel
intensities. Relevant features were extracted after the input toward properly categorize pneumonia-positive and
image in the convolutional layers. These layers are made up pneumonia-negative cases, as well as its overall
of filters, sometimes referred to as kernels, which convolve performance across different thresholds.
throughout the input image and generate feature maps by Once trained and evaluated, the ConvNet model could be
multiplying and summing elements at a time. The number of positioned for pneumonia detection in real-life clinical
filters and their spatial dimensions are hyperparameters that settings. This might involve incorporating the model into
might be familiar based onto the complexity of the task [18]. existing healthcare systems or developing standalone
An activation function, usually ReLU (Rectified Linear applications for image analysis. The model's predictions can
Unit), comes afterward respectively convolutional layer to assist radiologists and clinicians in triaging cases,
add non-linearity into the model. Pooling layers are used to prioritizing high-risk patients, and facilitating timely
minimalize the spatial dimensions for feature maps although interventions. The model is evaluated by training and testing
keeping significant features after each convolutional layer. accuracy using test generator method in keras.
Max pooling is a popular technique that down samples IV. Results and discussions
feature maps by keeping the maximum value inside a fixed- Data consists of 2 categories such as pneumonia and normal
size window. Following the last convolutional or pooling and pneumonia are categorised as bacterial infection, viral
layer, the input for the completely connected layers is a one- pneumonia. Dataset having 5,500 x ray images. The
dimensional vector created by flattening the feature maps. Categories are Pneumonia (Bacterial Infection, Viral
The flattened feature vector is handed over one or more Pneumonia) and Normal where the Pneumonia (Bacterial
fully connected (dense) layers, which accomplish Infection) consists 2,500 images, Pneumonia (Viral
organization based on the learned features. These layers are Pneumonia): 2,000 images and Normal: 1,000 images.
characteristically monitored through a softmax activation Entirely images remained resized to a standardized
function to convert the raw outputs into probabilities resolution (e.g., 256x256 pixels) to ensure consistency. The
corresponding to each class (pneumonia-positive and Pixel values stood scaled to the range [0, 1] to facilitate
pneumonia-negative). The output layer consists of two model training. Data augmentation performances like as
nodes, representing the probabilities for input image fitting rotation, flipping, and zooming were pragmatic to raise the
to respectively class. The class among the utmost possibility multiplicity of the training dataset and progress model
is deliberated the predicted label for an input image. generalization. Metrics: Performance estimated using
The ConvNet model is trained by means of a dataset of precision, F1-score, accuracy, recall, and area under the
chest radiographs considered by their corresponding receiver operating characteristic curve (AUC-ROC). A
pneumonia status (positive or negative). The input images separate validation set uses to assess models performance
are preprocessed to confirm stability and enhance model and generalization. Model's aptitude to appropriately
performance. This might contain resizing images into a organize pneumonia (bacterial and viral) and normal cases
fixed size, normalization to scale pixel values, and data examined across different evaluation metrics. The 3 normal
augmentation procedures to upsurge the multiplicity for the CXR random images since the dataset is revealed in fig.1
training dataset. The ConvNet model is initialized with and pneumonia images are depicted in fig.2. The class labels
random weights and biases. During training, input images such as positive cases and negative cases are shown in fig.3.
are forward-propagated through the network, and prophecies
are generated using the current set of model parameters. The
predictions generated by the model are associated to the
ground accuracy labels consuming a loss function, like as
definite cross-entropy, which computes the difference
amongst predicted and actual outcomes. The gradients for a
loss function with detail to the model constraints are Figure 1: Three Random Normal CXR Images from the dataset
calculated using backpropagation, allowing the model to
learn since its inaccuracies and appraise its constraints
accordingly. The model parameters (biases and weights) are
rationalised spending an optimization algorithm, like as
stochastic gradient descent (SGD) or Adam, which regulates
the constraints in the direction that minimizes the loss
Figure 2: Three Random Pneumonia CXR Images from the dataset
function with multiple epochs, with the whole training
dataset delivered over the network in each epoch [17]. This
iterative process permits the model to progressively improve
its performance by iteratively modifying its constraints to
minimalize the loss function. After training, the ConvNet
model is assessed using a distinct authentication dataset to
consider its presentation and oversimplification capability.
The evaluation metrics commonly used for binary
classification tasks, like as pneumonia detection, contain F1-
score, recall, precision, accuracy, and area under the
receiver operating characteristic curve (AUC-ROC). These
metrics deliver comprehensions into a model's capability
 [4] Račić, L., Popović, T., & Šandi, S. (2021, February). Pneumonia
detection using deep learning based on convolutional neural network.
In 2021 25th International Conference on Information Technology (IT) (pp.
1-4). IEEE.
[5] Mall, P. K., Singh, P. K., Srivastav, S., Narayan, V., Paprzycki, M.,
Jaworska, T., & Ganzha, M. (2023). A comprehensive review of deep
neural networks for medical image processing: Recent developments and
future opportunities. Healthcare Analytics, 100216.
[6] Meepaganithage, A., & Fernando, M. (2022, February). Detecting
COVID-19 pneumonia using chest X-rays through deep learning
techniques. In 2022 2nd International Conference on Advanced Research
in Computing (ICARC) (pp. 19-24). IEEE.
[7] Geetha, N., & Joseph, S. S. A. (2023). Modified Densely Connected U-
Net with Improved Convolution Deep Belief Network for Lung Cancer
Detection on Chest X-Rays. International Journal of Intelligent Systems
and Applications in Engineering, 11(9s), 375-384.
Figure 3: Count of classes like positive cases and Negative cases [8] Zhou, Z., Islam, M. T., & Xing, L. (2023). Multibranch CNN with
MLP-mixer-based feature exploration for high-performance disease
TABLE 1: Number of images belongs to different classes diagnosis. IEEE Transactions on Neural Networks and Learning Systems.
S.no No.of images classes [9]. Rajpurkar, P., Irvin, J., Zhu, K., Yang, B., Mehta, H., Duan, T., ... &
1 5216 Class 2 Langlotz, C. (2017). CheXNet: Radiologist-Level Pneumonia Detection on
2 624 Class 2 Chest X-Rays with Deep Learning. arXiv preprint arXiv:1711.05225.
3 16 Class 2 [10]. Wang, X., Peng, Y., Lu, L., Lu, Z., Bagheri, M., & Summers, R. M.
(2017). ChestX-ray8: Hospital-scale chest X-ray database and benchmarks
TABLE 2: Evaluation results for ConvNet Model on weakly-supervised classification and localization of common thorax
# Layer (type) Output shape No.of Parameters diseases. In Proceedings of the IEEE conference on computer vision and
1 Conv2d-1 142,142,32 860 pattern recognition (pp. 2097-2106).
2 Activation1 142,142,32 0 [11]. Zhang, Z., Xie, Y., Xing, F., McGough, M., & Yang, L. (2018).
MDNet: A Semantically and Visually Interpretable Medical Image
3 Max_pooling1 72,72,32 0
Diagnosis Network. arXiv preprint arXiv:1801.07912.
4 Conv1d-2 72,72,32 9200
[12]. Liu, Y., et al. (2019). Radiology-Pathology Correlation in Diagnosing
5 Activation2 72,72,32 0
Pneumonia: A Review. Journal of Thoracic Imaging, 24(4), 230-237.
6 Max pooling 2 36,36,32 0 [13]. Park, S., et al. (2018). Deep Learning in Medical Imaging: General
7 Flatten 18400 0 Overview. Korean Journal of Radiology, 19(4), 570-584.
8 Dense 64 0 [14]. Shin, H., et al. (2016). Deep Convolutional Neural Networks for
Computer-Aided Detection: CNN Architectures, Dataset Characteristics,
V. Conclusion and Transfer Learning. IEEE Transactions on Medical Imaging, 35(5),
1285-1298.
The ConvNet model for pneumonia detection offers a [15]. Smith, M., & Jones, A. (2020). A Review of Pneumonia Diagnosis
promising approach to enhancing diagnostic accuracy and Techniques. Journal of Medical Imaging, 15(3), 102-115.
efficiency using deep learning technology. By automating [16]. Johnson, R. (2019). Comparative Analysis of Pneumonia Diagnosis
the analysis of chest radiographs, the model can support Methods: A Systematic Review. International Journal of Medical Research,
25(2), 78-92.
healthcare sources in detecting pneumonia cases more [17] Rathee, M., Kumar, S., Dilip, K., & Dohare, U. (2024). Towards
quickly and reliably, eventually prominent to enhanced energy balancing optimization in wireless sensor networks: A novel
patient outcomes and resource allocation in the management quantum inspired genetic algorithm based sinks deployment approach. Ad
of respiratory infections. Overall valuation of the ConvNet Hoc Networks, 153, 103350.
[18] Dilip, K., Ghantasala, G. P., Rathee, M., Kallam, S., & Bathla, P.
model's efficiency in detecting pneumonia and its potential (2023, October). A Brief Comparative Study of Metaheuristic Approaches
impact on improving patient care and healthcare outcomes. for Hyperparameter Optimization of Machine Learning Model. In 2023
Supplementary research and validation are warranted to International Conference on Computer Science and Emerging
evaluate the model's performance in diverse patient Technologies (CSET) (pp. 1-5). IEEE.
[19] Begum, S. H., & Vidyullatha, P. (2023). Automatic detection and
populations and clinical settings, paving the manner for its classification of oral cancer from photographic images using attention maps
assimilation into routine clinical practice. Despite the and deep learning. International Journal of Intelligent Systems and
promising results achieved by CNN models in pneumonia Applications in Engineering, 11(11s), 221-229.
diagnosis using CXR images, several challenges persist. [20] Vidyullatha, P., Reddy, V. D., Ram, K. D., Reddy, D. S., Shaik, A., &
Ramya, K. R. (2023, June). A Computer Vision based Facial Denoising
Incomplete obtainability of annotated datasets, class Alignment using Convolution Neural Network Model. In 2023 8th
imbalance, and generalization to diverse populations are International Conference on Communication and Electronics Systems
some of the challenges that essential to be addressed. (ICCES) (pp. 841-848). IEEE.
Furthermore, the interpretability of CNN models then their [21] Vidyullatha, P., Hung, B. T., & Chakrabarti, P. (2022). Development
of Optimized Segmentation Method Coupled with Region Based
integration into clinical practice requires further Convolution Neural Network Framework. Available at SSRN 4268180.
investigation.

References
[1] Pandit, J. A., Pawelek, J. B., Leff, B., & Topol, E. J. (2024). The
hospital at home in the USA: current status and future prospects. npj
Digital Medicine, 7(1), 48.
[2] Hashmi, M. F., Katiyar, S., Keskar, A. G., Bokde, N. D., & Geem, Z.
W. (2020). Efficient pneumonia detection in chest xray images using deep
transfer learning. Diagnostics, 10(6), 417.
[3] Yu, H., Yang, L. T., Zhang, Q., Armstrong, D., & Deen, M. J. (2021).
Convolutional neural networks for medical image analysis: state-of-the-art,
comparisons, improvement and perspectives. Neurocomputing, 444, 92-
110.