Evaluating the Explainability of Vision Transformers in Medical Imaging

2025-10-13

AI Summary

Vision Transformers (ViTs) suffer from limited interpretability in medical imaging, hindering clinical trust and adoption. Method: We systematically evaluate the impact of architecture (ViT, DeiT, DINO, Swin Transformer) and pretraining strategy on explanation quality for blood cell and breast ultrasound classification tasks, and quantitatively and qualitatively compare Gradient Attention Rollout with Grad-CAM. Contribution/Results: DINO-pretrained models paired with Grad-CAM yield heatmaps exhibiting superior class discriminability and spatial localization accuracy—even for misclassified samples—consistently highlighting clinically relevant morphological features. This combination significantly outperforms alternatives across multiple datasets, establishing a verifiable, optimization-guided pathway toward trustworthy ViT deployment in clinical AI.

Technology Category

• Computer Vision: Interpretability, Explainability, and Transparency
• Machine Learning: Transparent, Interpretable, Explainable ML
• Humans and AI: Explainable AI (XAI) for Human Understanding

Application Category

• Security and Privacy: Data transparency and provenance
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Explainable and interpretable methods for personalization

Abstract

Understanding model decisions is crucial in medical imaging, where interpretability directly impacts clinical trust and adoption. Vision Transformers (ViTs) have demonstrated state-of-the-art performance in diagnostic imaging; however, their complex attention mechanisms pose challenges to explainability. This study evaluates the explainability of different Vision Transformer architectures and pre-training strategies - ViT, DeiT, DINO, and Swin Transformer - using Gradient Attention Rollout and Grad-CAM. We conduct both quantitative and qualitative analyses on two medical imaging tasks: peripheral blood cell classification and breast ultrasound image classification. Our findings indicate that DINO combined with Grad-CAM offers the most faithful and localized explanations across datasets. Grad-CAM consistently produces class-discriminative and spatially precise heatmaps, while Gradient Attention Rollout yields more scattered activations. Even in misclassification cases, DINO with Grad-CAM highlights clinically relevant morphological features that appear to have misled the model. By improving model transparency, this research supports the reliable and explainable integration of ViTs into critical medical diagnostic workflows.

Problem

• Evaluating Vision Transformers' explainability in medical imaging tasks
• Comparing explainability methods for different ViT architectures and strategies
• Improving model transparency for reliable clinical diagnostic integration

Innovation

• Evaluated ViT explainability using Gradient Attention Rollout
• Applied Grad-CAM for class-discriminative heatmap generation
• DINO with Grad-CAM provided most faithful medical explanations