Brain Image Parcellation Using Fully Convolutional Network with Adaptively Selected Features from Brain Atlases

Brain image parcellation is an important data processing step in neuroscience. Since multi-atlas based parcellation (MAP) uses prior information from brain atlases (i.e., manually labeled brainregions), it can provide accurate brain parcellation and has been widely adopted. Recently, some deep learning based brain image parcellation (DLP) methods using fully convolutional network (FCN) have been proposed. Compared with MAP, DLP has high computational efficiency, making it more applicable in practice. However, existing DLP methods either neglect or partially utilize brain atlases, making it difficult to get comparable parcellation accuracy as MAP. In this paper, we propose a new DLP method which is able to use brain atlases in an effective way. The network is based on FCN and non-local block based channel attention module (NL module). The input of our network is the target brain image to be parcellated as well as available brain atlases, and the parcellation result is produced through the FCN guided by the features of brain atlases selected by NL modules at different scales. In the experiments using two public MR brain image datasets (LPBA40 and NIREP-NA0), our method outperforms MAP and the state-of-the-art DLP methods due to the effective usage of brain atlases.