ENHANCING PNEUMONIA DETECTION WITH EXPLAINABLE AI THROUGH A DEEP LEARNING FRAMEWORK FOR CHEST RADIOGRAPH ANALYSIS
Abstract
Pneumonia is the leading infectious cause of death in children under five years of age, and its burden falls most heavily on low- and middle-income countries. Chest radiography is the first-line imaging test for diagnosis, but interpretation is time-consuming, varies between readers and is constrained by a shortage of radiologists. Convolutional neural networks (CNNs) can classify radiographs at expert level, yet their opaque, "black-box" nature remains a barrier to clinical trust and safe deployment. This study proposes an explainable deep learning framework for detecting pneumonia on chest radiographs. The framework combines Contrast Limited Adaptive Histogram Equalization (CLAHE) preprocessing and anatomically plausible data augmentation; a transfer-learned DenseNet-121 feature extractor and classifier fine-tuned with a class-weighted loss. The framework was evaluated on the public chest X-ray dataset (5,856 anteroposterior radiographs) and compared with VGG-16, ResNet-50 and EfficientNet-B0 under an identical protocol. On the held-out test set of 879 images, the proposed model achieved an accuracy of 95.0%, a precision of 95.0%, a sensitivity (recall) of 97.2%, a specificity of 91.5%, an F1-score of 96.1% and an area under the ROC curve (AUC) of 0.983, outperforming all three baselines. These results show that strong diagnostic performance and clinically verifiable explanations can be achieved within a single framework, supporting explainable AI as a second reader for chest radiographs, particularly in resource-limited settings.
Keywords
- pneumonia, chest X-ray, deep learning, explainable AI, Grad-CAM, LIME, SHAP, transfer learning, DenseNet-121, computer-aided diagnosis
Views & Downloads Statistics (Last 12 Months)
Monthly performance and trends of article views and downloads
