Abstract
Nuclei instance segmentation underpins quantitative pathology, yet its accuracy currently comes at a large parameter cost, and both the encoder and the per-task decoders must shrink where compute or latency is constrained. We present NucShare, centred on a single U-Net-style decoder shared by the nuclear pixel, horizontal-vertical, and classification tasks, with only a lightweight head per task. Sharing matches or slightly exceeds three independent decoders using 37% of their parameters. With a UNI2-h encoder it forms the teacher NucShare-T, with 690.5M parameters; the student NucShare-S keeps the decoder and adopts an off-the-shelf ConvNeXt V2-Tiny encoder, totalling 34.7M, transferring knowledge through output-level distillation. On PanNuke, NucShare-S retains 99.1% of the teacher's mPQ at 0.519, above CFR-SAM-H's 0.511; it runs 27.5x faster per tile than the teacher and reaches a zero-shot bPQ of 0.668 on MoNuSeg, above CFR-SAM-H's 0.662. Code: https://github.com/L-iMeng-L/Newidea.