Segment anything in pathology images with natural language.
Where this comes from
- Record sourced from PubMed, PMID 42722894.
- Also identified by DOI 10.1038/s43588-026-01042-5.
- No licence information is recorded for this record.
- Because redistribution is not established, this page shows the abstract only. Follow the links below for the full text.
Abstract
Segmenting tissues and cells in pathology images enables quantitative analysis but usually requires task-specific models or repeated spatial prompts. Here we show PathSegmentor, a foundation model that uses natural language descriptions to segment structures across anatomical regions and spatial scales. We assembled PathSeg from 21 public datasets, comprising 275,200 image-mask-label triples organized into a 3-level hierarchy of anatomical region, histological structure and object type. A single PathSegmentor model achieved the highest overall performance across 16 internal datasets and generalized to external public and clinical cohorts. Its text prompts reduced the need to identify every object with points or boxes and remained robust to variations in wording. We further used its predicted structures to explain breast cancer classification models through object-level perturbation and activation maps. These results establish a unified framework for flexible pathology segmentation with potential utility for clinically interpretable image analysis.