3D prostate MR image segmentation: A multi-task approach

Multi-atlas based approaches are effective for the medical image segmentation. The strategy of assigning weights for the atlases is critically important to the segmentation performance. Previous works either assign weights on the image level or assign weights of different regions independently, i.e., they can't employ the uniqueness of each region and the connectivity among different regions simultaneously. In this paper, a multi-task approach is proposed to reduce this drawback. To exploit the unique characteristic of each region, learning the segmentation result for each region is viewed as a single task. The weighted voting decision for each regions are made individually. To model the connectivity among different regions or tasks, a norm regularization term is introduced to refine the segmentation results made by each individual tasks. By this way, the proposed approach simultaneously exploits the unique character of each region and the connectivity among them. The proposed approach is tested on 60 3D prostate magnetic resonance (MR) images from 60 patients. Experiment results show that the proposed approach is comparative to or even superior to the state-of-the-art approaches for the prostate segmentation.