Comparative Analysis of Interpretable Deep Learning Models for Pneumonia Detection in Chest X-rays

Abstract

Pneumonia, a life-threatening respiratory infection caused by bacteria, viruses, or fungi, claims approximately 2.5 million lives annually, including 672,000 children under five. Misdiagnosis rates can exceed 20% in low-resource settings due to subjective chest X-ray interpretation. Our master’s thesis addresses this challenge with an effective computer-aided diagnosis (CAD) system using deep learning. Our model, a Convolutional Neural Network (CNN) with a Squeeze-and-Excitation (SE) attention layer, was evaluated alongside transfer learning models like DenseNet121 and ResNet50 but trained from scratch for resource efficiency in binary (pneumonia vs. normal) and multi-class (normal, pneumonia, other lung diseases) classification. Our comprehensive preprocessing pipeline, combining CLAHE for image enhancement with augmentations (e.g., rotation, flipping) and weighted loss, tackles class imbalance and boosts training robustness. Evaluated on the Kaggle Pneumonia Chest X-ray and NIH ChestX-ray14 datasets, our model achieves 97.5% accuracy and 0.9798 F1-score for binary classification, and 74.9% accuracy with 0.9660 F1-score for multi-class classification, with low loss and fast training times ideal for resource-constrained settings. Our model’s explainability, enhanced by Grad-CAM++ and SHAP, provides visual and feature-level insights, aligning with clinical needs for interpretable diagnostics.