Disentanglement-Based Unsupervised Domain Adaptation for Nuclei Instance Segmentation

Unsupervised domain adaptation (UDA) for nuclei instance segmentation is vital for pathology cell analysis. Though deep learning-based models have achieved promising performance on the nuclei instance segmentation task, they require rich and reliable annotations which are generally not available in digital pathology studies. UDA techniques can therefore address the domain difference among different modalities and limited annotation problems by transferring knowledge among domains. One key limitation of conventional UDA methods is the entanglement of domain-specific factors with the learned features during adaptation. To tackle this problem, we propose a UDA framework from the perspective of feature disentanglement for instance segmentation in microscopy images. More specifically, we design a novel multi-modal reconstruction (MMR) module to separate domain-invariant and domain-specific features with the least missing information via a feature reconstruction technique. Experimental results on two UDA instance segmentation tasks demonstrate that our proposed method outperforms the state-of-the-art UDA methods.