Supervise Less, See More: Training-free Nuclear Instance Segmentation with Prototype-Guided Prompting

TL;DR

SPROUT tackles training-free nuclear instance segmentation by uniting histology-informed priors with prototype-guided prompting to SAM. The method constructs slide-specific foreground/background prototypes from optical-density-based stain priors, aligns image features to these prototypes via a progressive partial optimal transport (POT-Scan) scheme, and converts the resulting activations into positive/negative prompts for SAM without any parameter updates. A containment-aware, soft-NMS refinement then yields accurate instance masks. Across MoNuSeg, CPM17, TNBC, and PanNuke, SPROUT demonstrates competitive performance relative to SAM-based, fully supervised, and weakly supervised baselines, while offering significant efficiency and strong robustness to hyperparameters, illustrating a practical zero-shot pathway for pathology-aware nuclear segmentation.

Abstract

Accurate nuclear instance segmentation is a pivotal task in computational pathology, supporting data-driven clinical insights and facilitating downstream translational applications. While large vision foundation models have shown promise for zero-shot biomedical segmentation, most existing approaches still depend on dense supervision and computationally expensive fine-tuning. Consequently, training-free methods present a compelling research direction, yet remain largely unexplored. In this work, we introduce SPROUT, a fully training- and annotation-free prompting framework for nuclear instance segmentation. SPROUT leverages histology-informed priors to construct slide-specific reference prototypes that mitigate domain gaps. These prototypes progressively guide feature alignment through a partial optimal transport scheme. The resulting foreground and background features are transformed into positive and negative point prompts, enabling the Segment Anything Model (SAM) to produce precise nuclear delineations without any parameter updates. Extensive experiments across multiple histopathology benchmarks demonstrate that SPROUT achieves competitive performance without supervision or retraining, establishing a novel paradigm for scalable, training-free nuclear instance segmentation in pathology.

Figures (28)

• Figure 1: Comparison of one-shot and proposed self-reference strategies in feature extraction. One-shot (blue box) fails to capture precise and diverse nuclei, even with similar pairs or backbones trained on natural and pathology images. Instead, our self-reference approach (red box) leverages high-confidence regions within the image to extract more robust features for similarity guidance.
• Figure 2: SPROUT pipeline for point prompt generation. It consists of three steps: (i) Feature–prototype similarity mapping: H&E stain priors are used to identify high-confidence foreground and background regions, from which clustering extracts representative prototypes that serve as anchors for similarity matching; (ii) POT-Scan: a partial optimal transport scheme that progressively aligns features to prototypes, filtering ambiguous assignments through partial mass transport; (iii) Activation prompting: prototype-reweighted activations are aggregated into foreground maps, from which positive and negative point prompts are sampled to guide SAM-based instance prediction.
• Figure 3: Performance comparison across supervision types. SPROUT consistently outperforms SAM-based (MedSAM), fully supervised (U-Net), and point-supervised (SC-Net) models across datasets with superior effectiveness in segmentation.
• Figure 4: Visualization of instance segmentation results from different methods. SPROUT delivers more correct instances with fewer overlaps. The highlighted regions show distinct differences.
• Figure 5: Performance comparison of pathology-based and natural-image-based backbones. On MoNuSeg (a, b) and CPM17 (c, d), the self-reference mask strategy mitigates the domain gap and achieves the best AJI at patch size $128\times128$ matching nuclear scale.

Theorems & Definitions (2)

• proof
• proof