Dense-CBAM-Forestnet: An Efficient Deep Learning Framework for Multi-Label Chest X-Ray Classification

Abstract

: Timely and reliable interpretation of chest X-ray (CXR) images remains a bottleneck in large-scale screening programmes, especially in regions where expert radiologists are scarce. This study proposes a modular three-stage framework, named Dense-CBAM-Forestnet (CBAM: Convolutional Block Attention Module), which couples contrast-adaptive preprocessing, a DenseNet-based feature extractor, and a lightweight attention-driven classification head to address the multi-label nature and strong class imbalance of CXR disease recognition. The network is trained end-to-end with class-balanced binary-cross-entropy loss on 88,000 images from the publicly available ChestX-ray14 corpus and optimised with cosine-annealed learning-rate scheduling. Quantitative evaluation on a held-out test split shows a mean Area Under Curve - Receiver Operating Characteristic (AUC-ROC) of 0.826, while the macro-average F1 score steadily rises during training and stabilises around 0.23. Ablation confirms that each stage contributes additively to sensitivity on rare pathologies. These findings indicate that the proposed pipeline, Dens-CBAM-Forestnet, delivers competitive diagnostic accuracy without inflating computational cost, suggesting its suitability for point-of-care triage and computer-aided reporting systems.