Cell-CAEW: Cell Instance Segmentation Based on ConvAttention and Enhanced Watershed

Cell instance segmentation in microscopy images is a challenging task. The morphological differences between different types of cells are significant, it is difficult to distinguish the boundaries between adjacent or overlapping cells. To address these issues, we improved Cellpose’s framework and proposed Cell-CoaT. Cell-CoaT adopts CoaT as the encoder and designs a decoder that can integrate features from different scales, and predicts the center region and gradient fields of cells. In the post-processing stage, we utilized a Marker-Controlled Watershed Segmentation with center point labels predicted by the network to alleviate under-segmentation and over-segmentation. Cell-CAEW obtains an F1 score of 0.7724 on the tuning set. The code will be released soon.