XAI-Annotate: Automated Medical Segmentation via Explainable AI Saliency Distillation

Rong Fu, Yaoyang WU, Anqi Zhao, weizhi Tang, Simon James Fong

Keywords: Explainable AI (XAI), Automatic Mask Generation, ​Annotation-Efficient Learning, ​Saliency-to-Mask Conversion, Breast Ultrasound Segmentation, ​Human-AI Collaboration, ​Computation-Efficient Annotation, Dice Coefficient

TL;DR: XAA transforms classification explainability into clinical annotations by distilling XAI heatmaps into segmentation masks, enabling radiologist-efficient refinement while preserving diagnostic fidelity.​

Abstract: The deployment of medical AI is hindered by the high cost of expert annotations, particularly for segmentation tasks. We introduce XAI-Annotate, a novel framework that transforms classification explainability into automated annotation. By leveraging saliency maps from correctly classified samples, our method distills latent segmentation cues into high-quality masks—without requiring segmentation labels. XAI-Annotate combines saliency-to-mask distillation, boundary-aware refinement with expert feedback, and multi-scale concordance validation to ensure clinical relevance. Validated on breast ultrasound data, our approach demonstrates significant annotation efficiency gains while maintaining strong segmentation performance. This work bridges the gap between interpretability and deployability in medical AI.

Primary Area: interpretability and explainable AI

Code Of Ethics: true

Submission Guidelines: true

Anonymous Url: true

No Acknowledgement Section: true

Submission Number: 1689