Multi-atlas Segmentation Combining Multi-task Local Label Learning and Semi-supervised Label Propagation.

Multi-atlas based segmentation methods have achieved great success in hippocampal segmentation due to their promising performance. However, the correlation between voxels in the target image is often ignored. In this study, an image segmentation method based on multi-atlas is proposed by combining the multi-task learning method with the semi-supervised label propagation algorithm. Both multi-task learning and semi-supervised label propagation can take advantage of the correlation between voxels in the target image. Specifically, instead of training an independent model for each voxel to be segmented in the target images, the multi-task learning method trains a joint classification model for a multi-task-voxel cluster in a target image patch, followed by majority voting to obtain a probabilistic segmentation for the central voxel. The probabilistic map is then used to guide a semi-supervised label propagation algorithm to get the final segmentation results. The proposed method was applied to hippocampus segmentation in MR images and compared with state-of-the-art multi-atlas segmentation methods. Experimental results demonstrated that our method had better segmentation performance for the hippocampal segmentation.