CLEAR: Consistency-augmented hierarchical concept bottlenecks for medical image classification.
Authors
Affiliations (7)
Affiliations (7)
- School of Information and Software Engineering, University of Electronic Science and Technology of China, Chengdu, Sichuan, 610054, China. Electronic address: [email protected].
- School of Information and Software Engineering, University of Electronic Science and Technology of China, Chengdu, Sichuan, 610054, China. Electronic address: [email protected].
- School of Information and Software Engineering, University of Electronic Science and Technology of China, Chengdu, Sichuan, 610054, China. Electronic address: [email protected].
- School of Information and Software Engineering, University of Electronic Science and Technology of China, Chengdu, Sichuan, 610054, China. Electronic address: [email protected].
- Department of Radiology, West China Hospital, Sichuan University, Chengdu, Sichuan, 610041, China. Electronic address: [email protected].
- College of Medical Engineering and Technology, Institute of Medical Engineering Interdisciplinary Research, Xinjiang Medical University, Ürümqi, Xinjiang, 830054, China. Electronic address: [email protected].
- School of Information and Software Engineering, University of Electronic Science and Technology of China, Chengdu, Sichuan, 610054, China; College of Medical Engineering and Technology, Institute of Medical Engineering Interdisciplinary Research, Xinjiang Medical University, Ürümqi, Xinjiang, 830054, China. Electronic address: [email protected].
Abstract
Concept Bottleneck Models (CBMs) enhance interpretability in medical image classification by introducing semantic concepts as intermediate representations. However, most existing CBMs adopt a flat concept bottleneck that predicts concepts of different semantic granularity from a shared visual representation, limiting the backbone's ability to learn concept-aligned features and leading to weak semantic-visual correspondence. Moreover, concept-level explanations alone are often insufficient in clinical practice, where intuitive and localized visual evidence is crucial. Although post-hoc heatmap-based explanations are commonly used as visual supplements, they are highly sensitive to intensity variations and input perturbations, resulting in unstable and unreliable interpretations in medical imaging. To address these challenges, we propose CLEAR, a Consistency-Augmented Hierarchical Concept Bottleneck framework for medical image classification that jointly enables hierarchical concept reasoning and robust visual explanation. CLEAR consists of a Progressively Guided Concept Perception Mechanism (PGCPM), which models local-to-global concept interactions in a structured manner inspired by clinical diagnostic reasoning, and a Multi-View Robust Heatmap Perception (MRHP) strategy, which aggregates patch-level importance across multiple intensity-perturbed views to produce stable and focused visual explanations. Extensive experiments on multiple medical imaging datasets demonstrate that CLEAR achieves competitive classification performance while providing more reliable and robust interpretability compared to existing CBM-based approaches.