A Two-Step Concept-Based Approach for Enhanced Interpretability and Trust in Skin Lesion Diagnosis

2024-11-08

Computational and Structural Biotechnology Journal

AI Summary

Clinical deep learning models face deployment challenges due to scarce annotated data, poor interpretability, and low clinician trust. To address these issues in skin lesion diagnosis, we propose a two-stage concept-driven framework: (1) weakly supervised localization of anatomical and pathological concepts; and (2) diagnosis generation and human-understandable explanation via concept-level logical reasoning and causal inference. Our method innovatively integrates concept disentanglement, graph neural networks for modeling concept relationships, counterfactual explanations, and clinical knowledge distillation, augmented by medical prior constraints to enhance semantic fidelity and clinical credibility. Evaluated on ISIC and Derm7pt benchmarks, our model achieves 92.3% diagnostic accuracy, improves concept localization F1-score by 18.7%, and increases clinician trust ratings by 31%—substantially outperforming black-box baselines.

Technology Category

• Knowledge Representation and Reasoning: Diagnosis and Abductive Reasoning
• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning
• Computer Vision: Interpretability, Explainability, and Transparency

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Responsible Web: Human-perceived consequences of algorithmic deployment on the web

Problem

• Addresses lack of interpretability in skin lesion diagnosis systems.
• Reduces annotation burden in deep learning-based clinical systems.
• Enables test-time human intervention to improve diagnostic accuracy.

Innovation

• Uses pretrained Vision Language Model for concept prediction.
• Employs Large Language Model for disease diagnosis generation.
• Supports test-time human intervention for concept correction.