Cross- and Intra-image Prototypical Learning for Multi-label Disease Diagnosis and Interpretation

• DOI: 10.48550/arxiv.2411.04607
• Publication Date: 2024-11-07
• Authors: Chong Wang; Fengbei Liu; Yuanhong Chen; Helen Frazer; Gustavo Carneiro
• Keywords: FOS: Computer and information sciences; Computer Vision and Pattern Recognition (cs.CV); Computer Science - Computer Vision and Pattern Recognition

Recent advances in prototypical learning have shown remarkable potential to provide useful decision interpretations associating activation maps and predictions with class-specific training prototypes. Such has been well-studied for various single-label diseases, but quite relevant more challenging multi-label diagnosis, where multiple diseases are often concurrent within an image, existing models struggle obtain meaningful effective class prototypes due the entanglement of diseases. In this paper, we present a novel Cross- Intra-image Prototypical Learning (CIPL) framework, accurate disease diagnosis interpretation from medical images. CIPL takes advantage common cross-image semantics disentangle when prototypes, allowing comprehensive understanding complicated pathological lesions. Furthermore, propose new two-level alignment-based regularisation strategy that effectively leverages consistent intra-image information enhance robustness predictive performance. Extensive experiments show our attains state-of-the-art (SOTA) classification accuracy two public benchmarks diagnosis: thoracic radiography fundus Quantitative interpretability results also superiority weakly-supervised localisation over other leading saliency- prototype-based explanation methods.