Segment anything in pathology images with natural language.

PathSegmentor, a foundation model using natural language, achieves top performance in pathology image segmentation across diverse datasets with 275,200 image-mask-label triples.

• Why it matters: Accurate tissue and cell segmentation in pathology images is crucial for quantitative analysis but often relies on task-specific models or extensive manual prompts, limiting flexibility and efficiency.
• What they did: The authors assembled PathSeg from multiple public datasets and trained a single model to interpret natural language prompts, enabling segmentation across anatomical regions, structures, and object types.
• The result: PathSegmentor outperformed existing methods, generalized well to external data, and facilitated explainable breast cancer classification, supporting more flexible and interpretable pathology image analysis.