SPROUT: Training-free Nuclear Instance Segmentation with Automatic Prompting

Wen Zhang, Qin Ren, Wenjing Liu, Haibin Ling, Chenyu You

Keywords: Nuclear instance segmentation; Optimal transport; Training-free methods; Digital pathology

TL;DR: SPROUT is a training-free framework that instantiates stain priors and prototype-guided optimal transport to prompt SAM, enabling robust and scalable nuclear segmentation.

Abstract:

Nuclear instance segmentation is a cornerstone task in digital pathology, with broad potential to drive clinical decision-making and accelerate therapeutic discovery. Recent advances in large vision foundation models have shown promise for zero-shot segmentation in biomedical domains. However, most efforts in pathology still rely on pre-trained vision models through fine-tuning or adapter modules. These approaches demand costly annotations and heavy computation, leaving efficient training-free methods largely unexplored.
To this end, we propose SPROUT, a simple yet effective framework for annotation-free prompting. Specifically, we leverage histology-informed stain priors to construct slide-specific references for mitigating domain gaps and instantiate a prototype-guided partial optimal transport scheme to progressively refine nuclear representations. In addition, we embed high-quality positive and negative prompts into the Segment Anything Model (SAM) without any fine-tuning.
Extensive experiments across multiple histopathology benchmark datasets demonstrate that SPROUT achieves competitive performance while requiring neither annotations nor retraining. These results establish SPROUT as a scalable, training-free solution for nuclear instance segmentation in computational pathology. Our codes are available at here.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 2165