MAUNet: Modality-Aware Anti-Ambiguity U-Net for Multi-Modality Cell Segmentation

Automatic cell segmentation enjoys great popularity with the development of deep learning. However, existing methods tend to focus on the binary segmentation between foreground and background in a single domain, but fail to generalize to multi-modality cell images and to exploit numerous valuable unlabeled data. To mitigate these limitations, we propose a Modality-aware Anti-ambiguity U-Net (MAUNet) in a unified deep model via an encoder-decoder structure for robust cell segmentation. The proposed MAUNet model enjoys several merits. First, the proposed instance-aware decode endows pixel features with better cell boundary discrimination capabilities benefiting from cell-wise distance field. And the ambiguity-aware decode aims at alleviating the domain gap caused by multimodality cell images credited to a customized anti-ambiguity proxy for domaininvariant learning. Second, we prepend the consistency regularization to enable exploration of unlabeled images, and a novel post-processing strategy to incorporate morphology prior to cell instance segmentation. Experimental results on the official validation set demonstrate the effectiveness of our method. Code and models are available at https://github.com/Woof6/neurips22-cellseg_saltfish .