A MULTI-ATLAS GUIDED 3D FULLY CONVOLUTIONAL NETWORK FOR MRI-BASED SUBCORTICAL SEGMENTATION

In this paper, we proposed and validated an effective multi atlas guided 3D fully convolutional network (FCN) for segmenting subcortical structures from magnetic resonance images (MRIs). In the multi-atlas framework, the label information and the image intensity information (in terms of a patch based manner) were introduced into the proposed network. To further learn both local and global contexts, we adopted long skip connection strategy to embed outputs from convolutional layers in deconvolutional layers, which encourages consistency between features extracted at different scales. In addition, a region limited sampling method was adopted to reduce the training time and improve the segmentation accuracy. Experiments were performed on a dataset consisting of 16 T1-weighted MRIs. Compared to several existing state-of-the-art segmentation methods for subcortical structures, including a multi-atlas joint label fusion method and a representative 3D FCN method, the proposed method performed significantly better for a majority of the subcortical structures.