Iterative multi-atlas based segmentation with multi-channel image registration and Jackknife Context Model

For medical image segmentation, multi-atlas based segmentation methods have attracted great attention recently. Within the multi-atlas segmentation framework, labels of all atlases are propagated to the target image by means of image registration and then fused to achieve segmentation of the target image. While most multi-atlas based segmentation methods focus on developing effective label fusion strategies, few of them make an effort to improve the accuracy of image registration between atlas and target images. Inspired by the idea that the estimated segmentation of the target image can be used to refine the pairwise registration performance, we propose an iterative strategy to improve registration accuracy between the atlas and target images using a multi-channel registration approach. In addition, an overfitting-resistant discriminative learning procedure, referred to as Jackknife Context Model (JCM), is adopted at each iteration to improve accuracy and robustness of label fusion results. Validation experiments on hippocampal segmentation have demonstrated that our method can statistically significantly improve the performance of the state-of-art multi-atlas based methods.