Patch-wise label propagation for MR brain segmentation based on multi-atlas images

In many neuroscience and clinical studies, accurate and automatic segmentation of subcortical structures is an important and difficult task. Multi-atlas-based segmentation method has been focus of considerable research due to its promising performance. In general, this technique first employs deformable image registration to construct the correspondences between pre-labeled atlas images and the target image. Then, using the acquired deformation field, labels in the atlas are propagated to the target image space. Obviously, anatomical differences between the target image and atlas images possibly affect the image registration accuracy, thus influencing the final segmentation performance. Another limitation is that the label propagation in most conventional multi-atlas based methods is implemented under a voxel-wise strategy, which cannot adequately utilize the local label information to determine the final label of the target sample. In this paper, we propose a patch-wise label propagation method based on multiple atlases for MR brain segmentation. First, each image patch is characterized by patch intensities and abundant texture features, to increase the accuracy of the patch-based similarity measurement. To determine the weights of the training patches for representing the test sample, a patch-based sparse coding procedure is employed. In the label propagation stage, to alleviate possible misalignment from the registration stage, we perform a patch-wise label propagation strategy in a nonlocal manner to predict the final label for each target sample. To evaluate the proposed segmentation method, we comprehensively implement our proposed method by conducting hippocampus segmentation on the ADNI data set. Experimental results demonstrate the effectiveness of the proposed method and show that the proposed method outperforms two conventional multi-atlas-based methods.