Structural Magnetic Resonance Imaging (MRI) is an effective tool for understanding the brain tissue and can differentiate between a neurotypical and a pathology affected brain. Brain segmentation of the different regions is often the first step in quantifying the extent of pathological infection. This important step, however, is difficult in developing fetal brains as a direct result of the relatively small volume of the brain and incomplete development. Manual segmentation is time consuming, cumbersome and prone to human errors. Hence, there is a crucial need to automate the segmentation process for the diagnosis of pathology and for potential intervention and treatment. In this paper, we study state-of-the-art learning based framework for multilabel atlas based segmentation, VoxelMorph on the FeTa 2022 dataset. Essentially, our work addresses the lack of standard brain volumes of pathologies by training a segmentation model only on neurotypical brains. We learn generalizable deformation parameters using the VoxelMorph architecture. We observe learning based atlas registration to achieve an average Dice score of 0.62 in the pathological FeTa 2022 MRI dataset, with an improvement of 0.07 over symmetric normalization based iterative atlas registration.
|