Distributionally Robust Segmentation of Abnormal Fetal Brain 3D MRI

The performance of deep neural networks typically increases with the number training images. However, not all images have same importance towards improved and robustness. In fetal brain MRI, abnormalities exacerbate variability developing anatomy compared to non-pathological cases. A small abnormal cases, as is available in clinical datasets used for training, are unlikely fairly represent rich brains. This leads machine learning systems trained by maximizing average be biased toward problem was recently referred hidden stratification. To suited use, automatic segmentation methods need reliably achieve high-quality outcomes also pathological this paper, we show that state-of-the-art pipeline nnU-Net has difficulties generalize unseen mitigate problem, propose train a network minimize percentile distribution per-volume loss over dataset. We can achieved using Distributionally Robust Optimization (DRO). DRO automatically reweights samples lower performance, encouraging perform more consistently on validated our approach dataset 368 T2w MRIs, including 124 MRIs open spina bifida cases 51 other severe development.