PM-CResUnet: A Lightweight Multi-Constraint Framework for Cell Segmentation and Counting

Abstract

Cell counting is a common task in the quantitative analysis of fluorescence microscopy images; however, variations in cell size, density, and boundary conditions often make the separation of adjacent cells difficult, thereby degrading counting accuracy. This paper proposes Pyramid-Multitask c-ResUnet (PM-CResUnet). Building upon the lightweight segmentation model c-ResUnet, we introduce pyramidal convolution and boundary peak weight maps, complemented by distance regression and instance-aware auxiliary supervision during training, and adopt complexity-aware adaptive augmentation (C3A) to expand the training data distribution. The model follows a segmentation-post-processing-counting pipeline, in which the auxiliary branches participate only during training and are excluded from the inference path. We systematically evaluate the contribution of each design choice to segmentation and counting through ablation studies on DSB2018. The results show that PM-CResUnet improves instance separation and counting performance while remaining lightweight, and that multi-scale features, boundary constraints, and data diversity act complementarily, with the full model achieving the best overall performance in both instance segmentation and counting.