GALOS: Generalizable Nuclei Segmentation via Lightweight Adaptation and Automatic Prompting

Segmenting nuclei in histopathology images is essential for disease diagnosis, but achieving robust generalization across varied staining protocols remains challenging. Although the Segment Anything Model (SAM) shows strong transfer learning on natural images, its reliance on manual prompts limits its applicability to dense nuclei segmentation. To address this, we present GALOS, a fully automated and domain-robust framework for nuclei segmentation. GALOS combines two core components: an Adaptive Gated LoRA Encoder (AG-Encoder) that efficiently adapts SAM's Vision Transformer using a parameterefficient gating mechanism, and an Otsu-Refinement Generator (OR-Gen) that generates coarse masks via Otsu thresholding and refines them through a lightweight projection layer. By eliminating the need for manual prompts, GALOS enables a truly automated segmentation workflow. Extensive evaluations on the MoNuSeg and TNBC datasets show that GALOS outperforms both traditional and SAM-based methods in segmentation accuracy and boundary precision, highlighting its strong cross-domain generalization. These results demonstrate that the synergy of AGEncoder and OR-Gen offers a high-fidelity, prompt-free solution for nuclei analysis in medical imaging.