Adaptable Global Network for Whole-Brain Segmentation with Symmetry Consistency Loss

Yuan-Xing Zhao,Yan-Ming Zhang,Ming Song,Cheng-Lin Liu

DOI: https://doi.org/10.1007/s12559-022-10011-9

IF: 4.89

2022-06-01

Cognitive Computation

Abstract:Segmenting the whole brain into a large number (for example, ) of regions is challenging due to the complexity of the brain and the lack of annotated data. Deep neural network-based segmentation methods have shown promise, but due to the limitation of graphics processing unit (GPU) memory, they cannot fully exploit the brain structure information contained in 3D data. This paper proposes a memory-efficient framework to exploit the global brain structure for whole-brain segmentation. In this framework, upon extracting the brain region by using a skull-stripping subnetwork, a global modeling subnetwork is used to learn a global brain representation for segmentation, while an adaptable segmentation subnetwork is used to optimize the global representation during training and directly segment the whole brain during testing. This framework enables the representation to be learned from the global structure with reduced memory consumption, and segmentation is performed without splitting the brain into patches. To overcome the lack of annotated data, we also propose a semi-supervised method based on a symmetry consistency loss and a prior knowledge-based pseudolabel generation strategy. Extensive experiments on four datasets demonstrate that our method outperforms previously developed methods and achieves state-of-the-art performance. The method is computationally efficient in that segmenting a raw magnetic resonance imaging (MRI) image requires less than 2 s on a TITAN X GPU; our approach is much faster than multiatlas-based methods and previously proposed 3D deep learning methods. The code is publicly available at https://github.com/ZYX-MLer/AGNetwork.

computer science, artificial intelligence,neurosciences