Evaluating Visual Explanations of Attention Maps for Transformer-Based Medical Imaging

2025-03-12

ISIC/iMIMIC/EARTH/DeCaF@MICCAI

AI Summary

This study addresses the limited clinical interpretability of Vision Transformers (ViTs) in medical imaging by systematically evaluating the clinical validity of attention maps for localizing critical pathological regions. We propose the first attention explanation evaluation framework tailored to medical imaging, introducing two novel quantitative metrics—*anatomical consistency* and *lesion sensitivity*—to assess attention map fidelity. Our methodology integrates multi-source validation: Grad-CAM and attention rollout visualizations, expert radiologist annotations, and statistical significance testing. Experiments on CheXpert and MIMIC-CXR reveal that only 38% of attention heatmaps achieve high spatial alignment with expert annotations, exposing substantial clinical bias in current ViT interpretability methods. This work quantifies the reliability boundary of Transformer-based attention explanations and establishes a reproducible, clinically grounded evaluation paradigm—thereby guiding future development of trustworthy, deployable explainable AI in radiology.

Technology Category

• Computer Vision: Interpretability, Explainability, and Transparency
• Machine Learning: Transparent, Interpretable, Explainable ML
• Knowledge Representation and Reasoning: Diagnosis and Abductive Reasoning

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Security and Privacy: Data transparency and provenance
• User Modeling, Personalization and Recommendation: Explainable and interpretable methods for personalization

Problem

• Evaluating attention maps for explainability in Vision Transformers.
• Comparing attention maps with other methods in medical imaging.
• Assessing context-dependent efficacy of attention maps in medical decisions.

Innovation

• Attention maps enhance ViT explainability in medical imaging.
• Comparison of attention maps with GradCAM and transformer-specific methods.
• Attention maps' efficacy varies, limited in comprehensive medical insights.