Integration Knowledge Distillation with Probabilistic-CAM Pooling for Detection and Classification 14 Types of Lung Disease from Chest X-Ray Images

Abstract

Accurate automated detection of thoracic diseases from chest X-ray images remains challenging due to subtle interclass differences and complex pathological patterns. This study proposes an interpretable CNN-CNN knowledge distillation framework that integrates Probability Class Activation Map (PCAM) Pooling to improve discriminative representation learning for single-label chest X-ray classification, using 91,324 single-label images (14 thoracic diseases plus a normal class, 15 classes total) from the NIH ChestX-Ray14 dataset. A pre-trained DenseNet-201 teacher guides a pre-trained DenseNet-169 student through a knowledge distillation strategy combining Binary Cross-Entropy loss and distillation loss. Unlike conventional post-hoc explainability techniques, PCAM is embedded directly into the embedding-generation process, enabling probabilityguided feature refinement without altering the original global average pooling structure. The distilled student model achieves a mean AUC of 0.809, higher than the teacher model's 0.787, demonstrating improved generalization, discriminative capability, and structurally consistent interpretability that support radiologist-assisted clinical validation.