MRIShift: Disentangled Representation Learning for 3D MRI Lesion Segmentation Under Distributional Shifts

In the clinical environment, heterogeneity of data coming from different centers poses challenges where the data may not conform to the assumption of being independent and identically distributed (i.i.d). As a result, it is essential to develop an approach that can generate better predictions even on samples it did not encounter during training. In this paper, we propose an unsupervised domain adaptation approach using disentangled representation learning to enhance the transferability of the 3D segmentation model for medical samples belonging to different distributions unseen during training. When dealing with medical image segmentation, it is essential to understand how features relate to voxel-wise labels. To achieve this, we focus on minimizing changes between data from different sources. This involves organizing the data into two categories: invariant components that remain consistent across different sources, and variant components that exhibit changes. By ensuring that domain shifts primarily affect the changing components, we aim to preserve the identifiability of latent variables. This helps in identifying joint data-label distributions within the target domain. We also introduce a new loss function that integrates image and feature level reconstruction with segmentation loss, enhancing the model's performance. Our experiments on real-world medical data demonstrate improved segmentation accuracy within known distributions and a notable 5% enhancement in handling out-of-distribution samples. Additionally, our method enhances the robustness of the model as demonstrated by the improvement in performance metrics such as R-AUC.