Concept Complement Bottleneck Model for Interpretable Medical Image Diagnosis

Hongmei WANG, Junlin Hou, Sunan He, Shu Yang, Hao Chen

Keywords: Explainable AI, medical image diagnosis, concept bottleneck model, concept discovery.

Abstract: Models based on human-understandable concepts have received extensive attention to improve model interpretability for trustworthy artificial intelligence in the field of medical image analysis. These methods can provide convincing explanations for model decisions but heavily rely on detailed annotations of predefined concepts. Consequently, they may be ineffective when concepts or annotations are incomplete or of low quality. Although some methods can automatically discover novel and effective visual concepts instead of relying on predefined ones, or generate human-understandable concepts using large language models, they often deviate from medical diagnostic evidence and remain difficult to interpret. In this paper, we propose a concept complement bottleneck model for interpretable medical image diagnosis. Specifically, we use cross-attention modules to extract key image features related to the predefined textual concepts, and employ independent concept adapters and bottleneck layers to distinguish concepts more effectively. Additionally, we devise a concept complement module to mine local concepts from the concept bank constructed using medical literature. The model jointly learns human-annotated predefined concepts and automatically discovered ones to improve performance in concept detection and disease diagnosis. Comprehensive experiments demonstrate that our model outperforms state-of-the-art methods while providing diverse and interpretable explanations.

Primary Subject Area: Interpretability and Explainable AI

Secondary Subject Area: Detection and Diagnosis

Registration Requirement: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 31