Semi-supervised Tissue Segmentation of 3D Brain MR Images

Clustering algorithms have been popularly applied in tissue segmentation in MRI. However, traditional clustering algorithms could not take advantage of some prior knowledge of data even when it does exist. In this paper, we propose a new approach to tissue segmentation of 3D brain MRI using semi-supervised spectral clustering. Spectral clustering algorithm is more powerful than traditional clustering algorithms since it models the voxel-to-voxel relationship as opposed to voxel-to-cluster relationships. In the semi-supervised spectral clustering, two types of instance-level constraints: must-link and cannot-link as background prior knowledge are incorporated into spectral clustering, and the self-tuning parameter is applied to avoid the selection of the scaling parameter of spectral clustering. The semi-supervised spectral clustering is an effective tissue segmentation method because of its advantages in (1) better discovery of real data structure since there is no cluster shape restriction, (2) high quality segmentation results as it can obtain the global optimal solutions in the relaxed continuous domain by eigen-decomposition and combines the pairwise constraints information. Experimental results on simulated and real MRI data demonstrate its effectiveness.