Multi-atlas subcortical segmentation: an orchestration of 3D fully convolutional network and generalized mixture function

To accurately segment subcortical structures and therefore profit for numerous neuroimaging applications, we proposed a multi-atlas subcortical segmentation method by orchestrating a 3D fully convolutional network and a generalized mixture function. Template atlases were first aligned to the target image. Then, target image patches and several most similar atlas patches were extracted from the transformed template atlases by employing a proposed similar atlas selection network and fed into the proposed multi-atlas driven 3D fully convolutional neural network. To sufficiently extract the subcortical features and improve the segmentation performance, a restricted region thought as a bounding box was utilized to roughly locate the subcortical structures. Additionally, a generalized mixture function was introduced to reduce the impact of the size and stride in 3D patch extraction. Two datasets consisting of 16 and 18 T1-weighted magnetic resonance images images (MRIs) were included to evaluate the proposed method, respectively. The results showed significantly higher segmentation accuracy than several state-of-the-art subcortical segmentation approaches for most subcortical structures. Furthermore, the proposed method achieved notable higher mean Dice similarity coefficients being, respectively, 0.915 and 0.869. The proposed method automatically and accurately segments subcortical structures in MRIs, which may assist the artificial diagnosis of brain disorders.