Outlining vestibular schwannoma (VS) on MRI is a routine but time-consuming step in radiotherapy planning and follow-up, and manual tracing differs from one observer to the next. Reported deep learning results are difficult to compare, because datasets, preprocessing, and evaluation protocols change from study to study. To separate architectural effects from experimental ones, we trained three representative networks -- a Plain 3D U-Net, an Attention U-Net with a Transformer bottleneck, and Swin UNETR -- on the public multi-center Vestibular-Schwannoma-MC-RC2 collection (185 patients, 534 MRI volumes). All three used the same patient-level split, preprocessing, optimizer settings, and evaluation protocol. Performance was measured with the Dice Similarity Coefficient (DSC), Intersection over Union (IoU), Precision, Recall, and the 95th-percentile Hausdorff Distance (HD95). The Attention U-Net with Transformer bottleneck was the best single model, with a test Dice of 0.8374 and an HD95 of 2.4190~mm. A validation-weighted soft-voting ensemble of the three networks raised Dice to 0.8408 while keeping boundary error low. These findings suggest that CNN-, attention-, and transformer-based models make complementary errors, and that weighted probability fusion is a practical way to obtain more robust VS segmentation on multicenter clinical MRI.

vestibular Schwannma,Medical Image Segmentation,3D U-net,Attention U-net,Swin UNETR