Semi-Supervised Sparse Label Fusion for Multi-atlas Based Segmentation

Multi-atlas based label fusion has shown great success in automatic and accurate medical image segmentation. However, most existing methods often equally and independently treat each voxel in labeling, and thus 1) find difficulty in discerning real and useful neighbors from those confusing ones in atlas images and 2) cannot use the structure information in images to be segmented. Motivated by these problems, in this paper, we propose a novel semi-supervised sparse method (SSSM) for multi-atlas label fusion. In the SSSM method, we first construct a unified graph with both labeled and unlabeled voxels and then use sparse representation to automatically determine the corresponding graph weights, which is followed by semi-supervised classification on the graph. Experimental results on segmenting brain anatomical structures in MR images show that our proposed method achieves not only improved accuracy but also more smooth segmented images, compared with conventional label fusion methods.