An Efficient Cross-Modal Segmentation Method for Vestibular Schwannoma and Cochlea on MRI Images

To obtain the segmentation results of vestibular schwannoma (VS) and cochlea on high-resolution T2 (hrT2) MR images according to the annotated contrast-enhanced T1 (ceT1) MR images, we propose an efficient cross-modal segmentation framework in this study. An image-to-image model is first applied to transfer ceT1 scans to hrT2 modality to alleviate the domain shift between them. In the model training phase, we adopted data augmentation for both original images and segmentation target regions to adapt to the diversity and heterogeneity of multi-center imaging data. Furthermore, random cropping along the z-axis and random flipping at all axial directions representing the different observation perspectives were implemented. We also utilized a 2.5D ResUnet model as the segmentation backbone. These strategies collectively contribute to improved segmentation output. Eventually, a post-processing method based on image contrast is applied to improve the quality of the pseudo-labels on the hrT2 modality. Our experimental results address the effectiveness of the proposed framework for the crossMoDA23 segmentation task of the vestibular schwannoma and cochlea on hrT2 modality, with average Dice scores of 0.8358 and 0.828 for VS and average Dice scores of 0.8355 and 0.844 for cochlea, respectively on validation and test data.