C.elegans Instance Segmentation via SAM Pseudo Labels and Morphology-Aware Tuning

Instance segmentation plays a crucial role in computer vision and biomedical image analysis. However, Caenorhabditis elegans (C. elegans) images pose challenges such as elongated shapes, dense overlaps, and partial occlusions, which limit the performance of traditional methods. To address these issues while reducing annotation costs, we propose a weakly supervised instance segmentation approach using pseudo labels generated by the Segment Anything Model (SAM), combined with morphology-aware optimization strategies. SAPNet is employed as the segmentation backbone, enhanced by three key improvements: an aspect ratio-based positive-negative sample selection mechanism, a partial occlusion-oriented negative sample augmentation strategy, and a loss weighting scheme based on object elongation. Experiments on the CSB-1 dataset show that our method achieves 41.9 AP using only point-level annotations, nearing the performance of fully supervised models. Ablation studies validate the independent effectiveness of each component. The results demonstrate strong robustness and generalization in complex biological imaging tasks, highlighting the practical value and broader application potential of the approach.