Go to MIDL 2026 Conference homepageUniNuc – Unified Automatic and Interactive Nucleus Instance Segmentation in Histopathology
Keywords: histopathology, instance segmentation, foundation model
Abstract: Accurate nucleus instance segmentation is a foundational task for computational pathology, yet dense cellularity, stain variability, and subtle boundaries make fully automatic pipelines prone to errors that must be efficiently corrected. Promptable segmentation foundation models such as the SAM provide an interface for human-in-the-loop refinement, but existing histopathology adaptations either prioritize fully automatic segmentation (often scaling to SAM-Huge for performance) or recover interactivity while retaining coarse decoding and bottom-up instance grouping that can fail in crowded tissue. We present \textbf{UniNuc}, a unified model that supports both automatic nucleus instance segmentation and iterative interactive refinement through a shared prompt interface. UniNuc (i) adopts an efficient SAM2 Hiera-B+ encoder together with a multi-scale high-quality mask decoder to preserve fine nuclear boundaries, and (ii) replaces heuristic pixel grouping with a DETR-style nuclei detector using a dedicated detection backbone whose predicted boxes serve as ``auto-prompts''. Optional language priors further improve nuclei type assignment. On PanNuke, UniNuc achieves $0.702$ bPQ and $0.529$ mPQ ($0.548$ mPQ with language priors), outperforming PromptNucSeg-H and CellViT-H while using substantially less compute than SAM-Huge-based baselines. On 14 wide-ranging datasets, UniNuc consistently improves interactive segmentation over PathoSAM-L in both in-domain and out-of-domain settings. Code and models will be publicly released.
Primary Subject Area: Segmentation
Secondary Subject Area: Application: Histopathology
Registration Requirement: Yes
Visa & Travel: Yes
Read CFP & Author Instructions: Yes
Originality Policy: Yes
Single-blind & Not Under Review Elsewhere: Yes
LLM Policy: Yes
Submission Number: 307
Loading