La présentation sera assurée par Vincent Jaouen et Guillaume Sallé :

Cross-modal segmentation for vestibular Schwanomma MR segmentation. Results on the CrossMoDa 2022 MICCAI Challenge Unsupervised domain adaptation (UDA) is studied to both mitigate 2 issues, often encountered in medical imaging : the domain shift and the lack of annotated data. In this work, we propose a new cross-modality UDA pipeline for vestibular schwanomma and cochlear segmentation. In the context of the CrossMoDA 2022 segmentation challenge, we propose to leverage a SinGAN model trained on a single axial slice to realistically blend synthetically altered versions of the regions of interest (ROI) for efficient data augmentation. Such a technique makes the training more robust by increasing the model's generalization ability. We show that this approach improves segmentation quality, reaching state of the art cross-modal segmentation performance and leading us to rank 3rd on the overall challenge and 1st on the main ROI segmentation task.