A Hybrid Deep Learning Approach for Pneumonia Detection from Chest X-Ray Images Using Custom CNN and Transfer Learning

Abstract

Pneumonia remains an important worldwide health concern, significantly affecting vulnerable populations such as children and the elderly. The timely and precise diagnosis of pneumonia using chest X-ray (CXR) imaging is essential; however, manual interpretation often faces challenges due to complex visual characteristics and inherent diagnostic subjectivity. This study introduces a hybrid deep learning framework that merges a customized Convolutional Neural Network (CNN) with the MobileNetV2 architecture, aimed at improving automated pneumonia classification in CXR images. The approach is structured into five primary phases: data preprocessing and increase, baseline CNN design, transfer learning utilizing VGG16 and
MobileNetV2, hybrid model integration, and thorough evaluation, including Grad-CAM visualizations. The dataset was meticulously separated into training, validation, and testing sets, employing augmentation techniques to enhance model robustness. Individual models, namely the standalone CNN, VGG16, and MobileNetV2, recorded accuracies of 87.98%, 90.06%, and
85.26%, respectively, while the hybrid model achieved a superior test accuracy of 90.06% with minimal loss. Notably, the hybrid architecture evidenced high F1-scores of 0.86 for NORMAL and 0.92 for PNEUMONIA classes, demonstrating an effective balance of precision and recall, further validated by Grad-CAM visualizations that illustrate clinically relevant areas of focus. Despite certain limitations such as dataset size and potential class imbalances, the proposed model effectively enhances pneumonia detection, positioning itself as a valuable asset for clinical decision-making, particularly in resource-limited surroundings.