Ask Before You Look: A conversational multi-level pathology segmentation framework

The development of AI co-pilots for digital pathology requires models that can understand clinical queries expressed in natural language and produce reliable segmentation masks. However, existing benchmarks generally assume that the queried target is present, leaving target-absent cases largely underexplored. Consequently, current language-guided segmentation models are not explicitly trained to determine whether a queried target exists and may hallucinate non-empty masks when it is absent. To address this limitation, we propose COMPASS, a COnversational Multi-level PAthology Segmentation System, which unifies target-presence verification and segmentation in a single end-to-end framework through an "Ask Before You Look" strategy. Target-presence verification is formulated as a supervised visual question answering objective and jointly optimized with the segmentation objective, allowing verification supervision to shape the shared representations used for segmentation. Beyond target verification, accurate pathology segmentation requires both fine-grained morphological information and broader tissue context. We therefore adopt pathology-oriented sparse multi-level feature fusion, where CKA-Spectral Anchor Selection selects complementary SAM-H encoder layers based on pathology-domain representational similarity. We further construct TumorVQA+, an explanation-augmented target-presence verification dataset, and ReasonPath, a pathologist-reviewed reasoning segmentation dataset. Experiments on tissue- and cell-level pathology benchmarks show that COMPASS improves language-guided pathology segmentation, reduces target-absent hallucination, and generates natural language explanations for segmented tumor regions.