FRAM‐Net: Feature Refinement Augmented Multi‐View Fusion Network for Chest X‐Ray Disease Classification

Abstract

Chest x‐ray (CXR) image classification is a representative multi‐label task in computer vision and medical image analysis, aiming to simultaneously identify multiple potential thoracic diseases from a single image to assist clinical diagnosis. However, CXR images often suffer from low grayscale contrast, significant structural variations across views, and extremely small lesion regions, which make it difficult for existing methods to learn feature representations that are both discriminative and robust. This challenge is further exacerbated in multi‐view scenarios, where semantic discrepancies between different imaging views increase the difficulty of accurate classification. To address these issues, we propose a Feature Refinement Augmented Multi‐view Fusion Network (FRAM‐Net). FRAM‐Net adopts a two‐stage architecture. In the first stage, an enhanced representation learning module is constructed based on information bottleneck theory to maximize the extraction of disease‐relevant information while suppressing redundant features. In the second stage, instance‐level contrastive learning is introduced to enforce semantic consistency across different views, and an adaptive multi‐view fusion mechanism with learnable weights is employed to effectively integrate complementary structural information from multiple views. Extensive experiments are conducted on the large‐scale public MIMIC‐CXR‐JPG dataset. The experimental results demonstrate that FRAM‐Net achieves the best mean AUC of 0.814, outperforming several advanced baseline methods by 0.4% to 1.9% in terms of overall classification performance.