FeST: Learning Spatio-Temporal Priors for Fetal Brain MRI Segmentation

Segmentation of anatomic brain structures on fetal magnetic resonance imaging is key in detecting and diagnosing congenital disorders. We propose FeST: a Fetal brain segmentation method, which includes information on the gestational age through Spatio-Temporal priors. We include gestational age in three different priors. We used it as input in our model through a sinusoidal encoding, and in the loss function through KL divergence and a size-prior for modeling the volumetric growth of the brain during development, and that anatomical structures of the brain grow at different rates [1]. We evaluate FeST in the FeTA dataset achieving a Dice similarity coefficient of 0.917, a Volume similarity of 0.974, and a 95th percentile Hausdorff distance of 10.96.