Contact-Aware Instance Geometry Prototypes with Dynamic Gram-Style Prompting for Source-Only Cross-Dataset Nucleus Segmentation

Abstract

Cross-dataset nucleus instance segmentation requires robust separation of touching nuclei despite changes in tissue appearance. We introduce contact-aware multi-granular instance geometry prototypes (CA-MIGP), a training-time objective that assigns scale-normalized radial states, separates exposed from contacting rims, and aligns local decoder features with cross-tissue state prototypes. An auxiliary Dynamic Gram-style prompt conditions the shared network on image appearance. We trained on PanNuke fold 1, selected settings on fold 2, and evaluated on held-out fold 3 and the retrospective external NuInsSeg and CryoNuSeg datasets, whose earlier outcomes informed method development. In a five-seed factorial study, adding CA-MIGP to the Gram-conditioned model increased equal-dataset image-mean panoptic quality by 0.0052 (95% paired-seed interval, 0.0005 to 0.0099); the dataset-specific changes were 0.0026, 0.0048, and 0.0083, respectively. Relative to a matched no-contact objective, NuInsSeg contact-pair success increased by 0.0236 across three seeds. The geometry branch is inactive at inference, leaving active parameters and FLOPs unchanged relative to the Gram-conditioned baseline. These findings support a conditional benefit of contact-aware prototype supervision within this Gram-conditioned architecture while leaving its deployment pathway unchanged. The external results are not independent unseen-domain validation. Code and reproducibility materials are publicly available.

The authors' abstract, as published at the source. Electronics, 2026 · DOI ↗