Accurate brain extraction on MRI using U-Net trained in two stages
Xue Li, Wenyao Zhang, Zijian Kang, Na Wang · 2022
Brain extraction is an essential processing step for most brain magnetic resonance imaging (MRI) studies. Due to the inaccuracy of available labels in training dataset, existing methods based on U-Net can only obtain rough brains. In this paper, we propose a new deep-learning-based method for accurate 3D brain extraction, in which a U-Net model is trained in two stages using different loss functions. In the first stage, the binary cross entropy (BCE) loss is used to train the model with original head MRIs and coarse labelled brain masks as usual U-Net models. In the second stage, a composite loss function that integrates active contour model (ACM) and BCE loss is introduced to guide the further training. By this means, the final trained model can not only strip head scalp and skull from head MRI scans, but also remove cerebrospinal fluid around brain tissues. Both quantitative and qualitative test results show that our brain extraction is more accurate than other counterparts. The improvement enables to build better brain model with more details.