Subcortical Brain Segmentation Based on a Novel Discriminative Dictionary Learning Method and Sparse Coding

Recently, many multi-atlas patch-based segmentation methods have been proposed and successfully implemented in various medical image applications. However, a precise segmentation of brain subcortical structures in a magnetic resonance image is still difficult since (1) brain MRI typically suffers low tissue contrast; and (2) image patterns around the boundary of a structure are similar such that similarity-based and reconstruction-based label fusion methods achieve inaccurate results. To overcome the above issues, we propose a novel discriminative dictionary learning method, which can simultaneously learn class-specific dictionaries and a shared dictionary from a set of brain atlases. In particular, we enforce a low-rank constraint on each class-specific dictionary, i.e. claim that its spanning subspace should have low-rank property. For the shared dictionary, a regularization term is used to minimize the between-class scatter of corresponding shared coefficients so that they can learn shared image patterns. The optimization algorithms are developed to solve the problems in the learning step. Under the multi-atlas patch-based segmentation framework, the whole learned dictionary then can be used for labeling the target image. The proposed low-rank discriminative dictionary and shared dictionary learning method has been evaluated on IBSR, LPBA40, and SATA MICCAI 2013 dataset for subcortical segmentation. The influence of different parameters was studied and the performance of the proposed method was also compared with the non-local patch-based segmentation, the sparse representation classifier based segmentation, the discriminative dictionary learning segmentation, and several deep learning methods. Experimental results establish the advantages of our method over these state-of-the-art methods.