Distilling Clinical Reasoning from Text Corpora for Explainable AI in Medical Imaging

Area of Science:

• Artificial Intelligence
• Medical Imaging Analysis
• Explainable AI (XAI)

Background:

• Deep learning in medical imaging lacks transparency, hindering clinical trust.
• Existing XAI methods provide limited clinical reasoning.
• Vision-language models often rely on superficial data correlations.

Purpose of the Study:

• To develop a novel teacher-student framework, K-Distill-XAI, for improved clinical reasoning in AI.
• To decouple visual feature learning from high-level clinical rationale generation.
• To enhance the trustworthiness and clinical utility of AI diagnostic tools.

Main Methods:

• A domain-expert Large Language Model (LLM) teacher was trained on biomedical literature.
• A multimodal student vision-language model was trained using cross-modal knowledge distillation.
• The student model was trained to align explanations with the teacher's expert reasoning.

Main Results:

• K-Distill-XAI significantly improved clinical accuracy over state-of-the-art baselines.
• Achieved an 8% relative improvement in CheXbert F1 score for medical report generation.
• Demonstrated state-of-the-art micro-averaged AUC across 14 clinical conditions.

Conclusions:

• K-Distill-XAI effectively generates clinically meaningful explanations, enhancing AI transparency.
• The proposed framework improves both diagnostic accuracy and classification performance.
• This approach offers a promising direction for trustworthy AI in healthcare.