| ABSTRACT |
Pneumonia remains a major global health concern, and chest X-ray (CXR) imaging is widely used to support its diagnosis. However, manual interpretation can be time-consuming and subject to interobserver variability, particularly when clinicians differ in experience. This study develops and evaluates a deep learning approach for automated pneumonia detection, incorporating explainable artificial intelligence to improve the interpretability of model predictions. Three architectures are implemented for binary classification of CXR images into normal and pneumonia categories: a custom convolutional neural network (CNN) trained from scratch as a baseline, and two transfer learning models, ResNet50 and EfficientNet, fine-tuned using pretrained feature representations. All models are trained and tested using a publicly available CXR dataset, and their classification performance is evaluated using accuracy, precision, recall (sensitivity), and F1-score. Gradient-weighted Class Activation Mapping (Grad-CAM) is applied to the best-performing model to generate heatmaps highlighting image regions that contribute most strongly to its predictions. The comparative evaluation examines the performance of the custom CNN and transfer learning approaches, while the heatmaps provide visual explanations of the model¿s classification decisions. This approach offers a basis for developing interpretable, automated tools to support pneumonia screening, with further clinical validation needed to establish their reliability and applicability in practice.
|
