Sparse Point-to-Pseudo-Instance Supervision for Annotation-Efficient Nuclei Segmentation

Accurate nuclei instance segmentation requires pixel-level boundary annotation, which is expensive and time-consuming in histopathology images. We propose a sparse point-to-pseudo-instance supervision framework that learns from limited center-point annotations. The framework first completes sparse nucleus centers through curriculum-based center recovery and then generates instance-level pseudo labels using point-prompted foundation-model proposals. For completed centers without valid proposals, point-derived fallback regions are introduced to recover missing supervision, and all generated regions are assembled into unified pseudo-instance labels. The generated labels are converted into the native training targets of different downstream segmentation models, including HoVerNet and StarDist, while inference relies only on the trained segmentation networks. Experiments on MoNuSeg, CPM17, and CoNSeP evaluate segmentation performance under limited point budgets and analyze the coverage–precision trade-off introduced by fallback recovery. Under a 10% point annotation budget, the framework achieves AJI and PQ scores of 0.6233 and 0.6036 on MoNuSeg and 0.6653 and 0.6626 on CPM17, respectively. Point budgets cover training; fully annotated validation subsets support model selection. Additional validation analyses quantify recovered nuclei and false additions, showing that fallback improves supervision coverage while its effectiveness depends on the proposal coverage and center accuracy. The results demonstrate that sparse point annotations can be effectively converted into transferable pseudo-instance supervision for nuclei instance segmentation.