Applications of Optimizing Neural Networks in Medical Image Registration

Abstract

We formulate feature-based medical image registration as a;point matching problem. Point matching requires us to solve for the (rigid or non-rigid) spatial mapping that brings one point-set into register with the other. In order to solve for the spatial mapping, we require the point-to-point correspondences between the two point-sets. Since this information is typically unavailable, we formulate the problem w.r.t. the spatial mapping and the correspondences as a mixed variable objective function. Deterministic annealing and the softassign are used to obtain a robust point matching algorithm. Results are shown for 2D rigid autoradiograph alignment and 3D non-rigid mapping of cortical anatomical MR brain structures.