"Why Is There a Tumor?": Tell Me the Reason, Show Me the Evidence.

Area of Science:

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

Background:

• Current medical AI models for tumor detection and segmentation often lack clinical semantic grounding.
• Existing models provide either localization (where) or justification (why), but not both, limiting clinical trust.
• There is a need for AI systems that can spatially localize findings and provide clinically relevant explanations.

Purpose of the Study:

• To develop AI models capable of justifying segmentation or detection using clinically relevant terms and pointing to visual evidence.
• To bridge the gap between AI-driven image analysis and clinical semantic understanding.
• To enhance the trustworthiness and interpretability of medical AI in clinical practice.

Main Methods:

• Curated a large-scale dataset of 180K image-mask-rationale triples, validated by expert radiologists.
• Designed rationale-informed optimization for self-supervised disentanglement and localization of clinical concepts.
• Developed models that integrate segmentation/detection with textual justifications and visual evidence grounding.

Main Results:

• The proposed model demonstrated superior performance in medical image segmentation and detection tasks.
• The model successfully grounds clinical concepts in spatially localized visual evidence.
• Experimental results across medical benchmarks validate the effectiveness of the rationale-informed approach.

Conclusions:

• The developed AI models can provide both localization and justification, improving clinical trust and utility.
• The rationale dataset and optimization strategy are effective for training interpretable medical AI.
• This work advances the field of explainable AI in medical imaging, paving the way for more reliable clinical decision support.