PathXAI improves pathology reports using tissue segmentation and AI
PathXAI: Evidence-guided structured pathology reporting by transferable tissue segmentation
Computer Vision and Pattern RecognitionComputation and Language
Summary
Pathology reports are crucial for diagnosing diseases by examining tissue samples under a microscope, but creating detailed and accurate reports can be challenging. The authors developed PathXAI, a system that uses advanced image analysis to identify and measure different tissue types in pathology slides, helping to produce clearer and more accurate diagnostic reports. This system showed better performance than other methods in matching expert diagnoses and describing tissue features. It also generates reports linked to visual evidence without needing to retrain its AI model.
What this means in practice
- •For clinical pathology labs: Generate structured pathology reports with improved accuracy by integrating tissue segmentation data and AI guidance.
- •For medical imaging software developers: Enhance pathology software with pretrained vision-language models that produce evidence-linked diagnostic reports without retraining.$Commercial implications: Supports development of advanced pathology reporting tools that combine tissue segmentation and AI, enabling commercial products with better diagnostic support.
Authors
Zhaorui Tan, Jun Aun Tan, Weixiang Zhong, Xiaoxi Wang, Tan Pan, Xingdong Xue, Li Liu, Jing Zhang, Jing Zhang, Weimiao Yu
Abstract
Pathology reporting requires translating tissue observations across regions and slides into specific diagnoses and supporting morphological descriptions. However, the use of explicit tissue-composition measurements to guide H&E-based report generation remains underexplored. Thus, we developed PathXAI, which combines segmentation-derived tissue information with H&E images and optional domain knowledge to guide diagnostic reasoning, targeted reinspection and structured report writing. Its segmentation component exceeded baselines in mean Dice on internal and external tests. In automated report evaluation on 1,200 public and 151 private cases, PathXAI achieved higher agreement with reference diagnoses and documented morphology than the H&E-based comparators. On the public evaluation cohort, PathXAI-Report exceeded Qwen + H&E in cancer family match rate by 4.4 percentage points and in finding F1 by 5.7 percentage points. PathXAI generates evidence-linked reports with a self-hosted pretrained vision-language model, without updating its weights.