Impact of Curvature on Intensity-Based Non-rigid Medical Image Registration

The non-rigid registration problem has gained much prominence in the area of medical image analysis. Graph cut-based solution to such a non-rigid registration problem is shown to improve the accuracy of the registration results over other existing approaches. Incorporation of curvature information in graph cut-based non-rigid bio-medical image registration has not been explored yet. In this work, a novel approach based on curvature is proposed to improve the performance of the Graph-Cut solution. In this method, first of its kind, the impact of curvature in an intensity-based non-rigid registration is well demonstrated. At first, the extraction of mean and Gaussian curvatures from the brain MRI are outlined and shown appropriate results. Based on the mismatch pattern in curvature distribution, a local displacement map (LDM) is generated for every such unmatched pixel in source image. The map essentially indicates the local displacement such as source pixel should undergo within a local neighborhood to achieve best curvature-wise matching with the corresponding target pixel. Experimental results clearly corroborate to the fact that intensity-based registration along with curvature can yield better registration accuracy.