OTLesMix boosts brain lesion segmentation by 6.6 Dice points

Data augmentation has become essential for training robust medical imaging segmentation models, but current mixing strategies limit the variability of generated lesion shapes and locations. Enter OTLesMix, a novel synthesis method from Robin Trombetta and Carole Lartizien, now on arXiv (2608.06264). It leverages a Wasserstein barycenter and an optimal transport plan to morph and combine real lesion patches into synthetic counterparts. This allows the model to see lesions that are not simple interpolations but geometrically and spatially diverse, expanding the effective training distribution while staying anatomically plausible.

The team evaluated OTLesMix on three brain lesion segmentation tasks (likely tumor, stroke, and white matter lesions). Compared to a model trained without synthetic data, OTLesMix boosted Dice scores by 2.9 to 6.6 points across tasks, and it also outperformed existing mix-based methods like CutMix and MixUp-based variants. The gains were most significant when training data was limited, a common bottleneck in medical AI. While this is a research preprint, the method could easily be integrated into existing segmentation pipelines, offering a practical way to improve model generalizability without requiring new annotated data.