MS-MT++: Enhanced Multi-scale Mean Teacher for Cross-Modality Vestibular Schwannoma and Cochlea Segmentation

Domain shift has been a long-standing issue for medical image segmentation. Unsupervised domain adaptation (UDA) methods have recently achieved promising cross-modality segmentation performance by distilling knowledge from a label-rich source domain to a target domain without labels. Different from CrossMoDA 2022, the challenge of 2023 includes highly heterogeneous MRI scans from more institutions and various scanners and subdivides the segmentation object into three key brain structures (intra/extra-vestibular schwannoma and cochlea), increasing the difficulty of domain adaptation. In this work, we improve our previous method and propose an enhanced multi-scale self-ensembling-based UDA framework for automatic segmentation of Vestibular Schwannoma and Cochlea on high-resolution T2 images. Our method demonstrates a mean Dice score of 0.669 and 0.766 for the extra/intra-VS joint area and Cochlea respectively, securing a top 5 finish in the CrossMoDA 2023 challenge.