The effectiveness of geometry features on multi-resolution diffeomorphic demons registration in the implementation of human cortex surface parcellation

Abstract

Fast automated labeling of the human cerebral cortex remains a challenging problem. We have implemented a completely automated pipeline to analyze brain morphology. For labeling of the cerebral cortex, a spherical diffeomorphic demons registration is used to drive an atlas surface into correspondence with the subject surface. This study focuses on identifying features or combination of features that will provide optimal correspondence between the atlas and subject. A multiresolution scheme is used for the surface registration procedure. The roles of anatomical features in multi-resolution surface registration have been evaluated by using them across all resolutions of the registration procedure as well as combining them from low to high anatomical detail corresponding to the registration refinement. We found that combining features from low to high anatomical features provided the best automated labeling of the surface. The mean Dice metric for the four lobar regions studied was 0.85 using the combined features.

Automated parcellation on the surface of human cerebral cortex generated from MR images

“…Table 2 and "DistanceToPCAP", contain more global information in showing polar points of the surface with extreme values, while "MeanCurvature" focuses more on local convexities, and "DistanceToHull" helps to identify overall deep fissures of the cortical surface. We did a study about the effectiveness of using these geometry features in surface registration [143], which led to the conclusion that since the geometry features have the properties of representing cortical surfaces from coarse to fine in the order of "DistanceToPCIS/AP", "DistanceToHull", and "Mean Curvature", if we apply them to the corresponding resolution levels of the multi-resolution registration, we can get better results of surface parcellation (please also see the detail in reference [143]).…”

“…2.50. We extend the histogram matching to the scalar values of spheres [143]. First, the histograms of the scalar values on the input sphere (the moving sphere) and on the reference sphere (the fixed sphere) are generated by setting the number of bins (256), the maximum value (1.0) and the minimum value (0.0) between which the histogram is going to be generated, as shown in Fig.…”