PathSegmentor is the first text-prompt-based baseline model for pathology image segmentation. To overcome the challenges of limited annotation data and category definition, we present PathSeg, a large-scale pathology segmentation dataset containing 275,000 image-mask-label triples collected from 21 publicly available sources. PathSegmentor allows users to perform semantic segmentation using natural language prompts, eliminating the need for spatial inputs such as points or boxes. Experimental results show that PathSegmentor is more accurate and has a wider range of applicability than existing spatial and text-prompt models, achieving an overall Dice score of 0.145 and 0.429, respectively. Furthermore, it enhances the interpretability of diagnostic models through feature importance estimation and imaging biomarker discovery, supporting pathologists' clinical decision-making.
