Unseen Domain Generalization for Prostate MRI Segmentation via Disentangled Representations

In clinical practice, medical images obtained from different sites often exhibit appearance variations, resulting in limited generalizability of deep learning models for segmentation in deployment. It is an important but challenging task to train a model which can directly generalize to unseen domains with distribution shifts. In this paper, we propose to disentangle content from style representations for prostate MRI segmentation to improve the model generalization, considering anatomical content information is domain invariant and decides the segmentation masks. Our method roots in a representation disentanglement network, sharing the content encoder with the segmentation module to remove the effect of appearance discrepancy. Besides, we introduce two domain discriminators to further regularize the disentangled representation learning. We extensively validate our model on a multi-site dataset for prostate MRI segmentation. Both quantitative and qualitative experimental results demonstrate the effectiveness of our method, outperforming the baseline method and many state-ofthe-art generalization methods.