Atlas-based registration parameters in segmenting sub-cortical regions from brain MRI-images

Multi-atlas segmentation has been proved to perform well in segmenting sub-cortical structures from images. In this work, we study different components of multi-atlas segmentation and propose new techniques to improve the segmentation accuracy. We found that the use of gradient information in addition to standard normalised mutual information increases the registration accuracy. We also studied different techniques to select atlases in the multi-atlas segmentation. In addition, the expectation maximisation algorithm was used to combine multi-atlas and intensity model information. The average similarity index obtained for six subcortical structures was 0.84.