干旱区生态环境知识资源中心

In this paper, we train a state-of-the-art deep neural network segmentation model to do fast brain volumetric segmentation from T1 MRI scans. We use image data from the ADNI and OASIS image collections and corresponding FreeSurfer automated segmentations to train our segmentation model. The model is able to do whole brain segmentation across 13 anatomical classes in seconds; in contrast, FreeSurfer takes several hours per volume. We show that this trained model can be used as a prior for other segmentation tasks, and that pre-training the model in this manner leads to better brain structure segmentation performance on a small dataset of expert-given manual segmentations.