Non-linear Cerebral Registration with Sulcal Constraints

In earlier work [1], we demonstrated that cortical registration could be improved on simulated data by using blurred, geometric, image-based features (Lvv) or explicitly extracted and blurred sulcal traces on simulated data. Here, the technique is modified to incorporate sulcal ribbons in conjunction with a chamfer distance objective function to improve registration in real MRI data as well. Experiments with 10 simulated data sets demonstrate a 56% reduction in residual sulcal registration error (from 3.4 to 1.5mm, on average) when compared to automatic linear registration and an 28% improvement over our previously published non-linear technique (from 2.1 to 1.5mm). The simulation results are confirmed by experiments with real MRI data from young normal subjects, where sulcal misregistration is reduced by 20% (from 5.0mm to 4.0mm) and 11% (from 4.5 to 4.0mm) over the standard linear and nonlinear registration methods, respectively.