Bootstrapping fully-automatic temporal fetal brain segmentation in volumetric MRI time series

Abstract

We present a method for bootstrapping training data for the task of segmenting fetal brains in volumetric MRI time series data. Temporal analysis of MRI images requires accurate segmentation across frames, despite large amounts of unpredictable motion. We use the predicted segmentations of a baseline model and leverage anatomical structure of the fetal brain to automatically select the "good frames" that have accurate segmentations. We use these good frames to bootstrap further model training. We also introduce a novel temporal segmentation model that predicts segmentations using a history of previous segmentations, thus utilizing the temporal nature of the data. Our results show that these two approaches do not provide conclusive improvements to the quality of segmentations. Further exploration into the automatic choice of good frames is needed before