Multi-scale and Focal Region Based Deep Learning Network for Fine Brain Parcellation

Brain parcellation plays an important role in neurodegenerative disease diagnosis and brain network analysis. One of the big challenges in brain parcellation is lack of clear anatomical boundary between different brain regions. As a result, for the task involving a large number of brain regions, i.e., during fine brain parcellation, the parcellation accuracy could be significantly degraded. Unfortunately, few studies focused on this issue. To this end, we propose a novel multi-scale deep brain parcellation network. Specifically, different scales of brain regions, i.e., global, middle and fine scales, are defined. From global to fine scales, brain regions are gradually subdivided and refined. The proposed deep network performs brain parcellation at each scale simultaneously (multi-task), where parcellation at fine scale is under the constraint of large scales. In addition, we also present a new focal region based auxiliary network, which focuses on the brain regions difficult to be parcellated at fine scale. The final parcellation results are obtained by merging the outputs of the brain parcellation backbone at all scales and the focal region based auxiliary network. Comparison and ablation experiments are conducted on a multi-center clinical brain MRI dataset of 267 subjects with 101 brain regions. Experimental results demonstrate that the proposed approach outperforms the state-of-the-art methods under comparison.