Joint Neighboring Coding with a Low-Rank Constraint for Multi-Atlas Based Image Segmentation

Multi-atlas methods have been successful for solving many medical image segmentation problems. Under the multi-atlas segmentation framework, labels of atlases are first propagated to the target image space with the deformation fields generated by registering atlas images onto a target image, and then these labels are fused to obtain the final segmentation. While many label fusion strategies have been developed, weighting based label fusion methods have attracted considerable attention. In this paper, we first present a unified framework for weighting based label fusion methods. Under this unified framework, we find that most of recent developed weighting based label fusion methods jointly consider the pair-wise dependency between atlases. However, they independently label voxels to be segmented, ignoring their neighboring spatial structure that might be informative for obtaining robust segmentation results for noisy images. Taking into consideration of potential correlation among neighboring voxels to be segmented, we propose a joint coding method (JCM) with a low-rank constraint for the multi-atlas based image segmentation in a general framework that unifies existing weighting based label fusion methods. The method has been validated for segmenting hippocampus from MR images. It is demonstrated that our method can achieve competitive segmentation performance as the state-of-the-art methods, especially when the quality of images is poor.