One Click per Cell Type Suffices: Training-free Group Interaction for Cell Instance Segmentation

This paper introduces Chain-of-Prompts, a training-free interactive framework built on the frozen image encoder of SAM3 that changes cell instance segmentation from one click per instance to one click per cell type. The key insight is that SAM3’s encoder already clusters same-type cells in feature space before prompting, so the method uses Hierarchical Similarity Gating to combine high-resolution and low-resolution cosine similarity maps, then applies non-parametric thresholding and connected-component labeling to extract reliable prompt points with precision above 96%. Farthest Prompt Recursion then selects the next prompt as the spatially farthest unreached reliable point, recursively expanding coverage without learned parameters. On three cell-type-annotated benchmarks—CoNIC, CoNSeP, and GlaS—CoP with only about 3 clicks per image retains over 90% of per-instance SAM3 performance while cutting annotation cost by 97% and even surpassing fully supervised models such as CellViT, CA-SAM2, and CellPose3. On four morphologically homogeneous benchmarks—MoNuSeg, TNBC, CryoNuSeg, and CPM-17—a single click per image preserves over 99% of per-instance prompting performance. The ablation study shows that removing recursive propagation drops AJI from 0.579 to 0.203, while the full method reaches 92.7% of the upper-bound AJI in an example with 3 clicks versus 245 per-instance clicks, demonstrating that group prompting is a practical way to scale foundation-model interaction to dense histopathology images.