One Sequence to Segment Them All: Efficient Data Augmentation for CT and MRI Cross-Domain 3D Spine Segmentation

Deep learning models for medical image segmentation often fail when applied to imaging protocols or modalities they weren't trained on—a major bottleneck for clinical deployment. This is especially acute for spine segmentation, where CT and MRI sequences vary widely. Researchers at TU Munich, Imperial College, and Université de Montréal tackled this with targeted data augmentation. They trained three spine segmentation models each on a single-modality/sequence dataset, then evaluated across seven out-of-distribution datasets spanning both CT and MRI. The results were striking: an average Dice score gain of 155% on unseen domains, with an in-domain accuracy degradation of only 0.008%—effectively zero. This demonstrates robust cross-modality transfer without sacrificing performance on the original training data.

To make strong augmentation practical, the team implemented GPU-optimized operations that not only avoided the usual slowdown but actually improved training efficiency by approximately 10%. They released their approach as an open-source toolbox, easily integrated into popular frameworks like nnUNet and MONAI. This work directly addresses the heterogeneity of clinical imaging, where models must handle new sequences or contrasts without expensive retraining. By providing a simple, computationally efficient augmentation pipeline, it enables more reliable spine segmentation in real-world diagnostic workflows, potentially reducing the need for extensive annotated multi-protocol datasets.