Watershed-enhanced cell instance segmentation based on guidance of gradient

Zhengwei Lu, Zehan Zhang

Abstract: Cell Segmentation in Multi-modality Microscopy Images NeurIPS 2022 competition aims to benchmark cell segmentation methods that could be applied to various microscopy images across multiple imaging platforms and tissue types. Due to the difficulty of obtaining labeled data, the competition team set this cell segmentation problem as a weakly supervised task and provided some labeled data and a large amount of unlabeled data. We constructed a network structure with CoaT as the encoder and designed a decoder for fusing different scale features, and conducted related experiments with full supervision and semi-supervision respectively. We followed cellpose's strategy and constructed a model to predict the central region of individual cells, and obtained the final instance segmentation results by the watershed algorithm. The method we proposed obtained an F1 score of 0.7724 on the tuning set.