Optimal data augmentation strategy for domain generalization in fetal brain MRI segmentation

Determine the optimal set and configuration of data augmentation techniques for training deep learning models to maximize cross-domain generalization in fetal brain MRI multi-class tissue segmentation, accounting for trade-offs between in-domain and out-of-domain performance.

Background

The paper finds that image augmentation is critical for improving generalization across sites, noting effective use of random bias fields, motion artifacts, and style/photometric augmentations by top teams. Despite these insights, the authors state that the optimal augmentation choices are not yet established.

They also discuss a potential trade-off between in-domain and out-of-domain generalization, which complicates selecting augmentation strategies that are universally beneficial.