Automatic Registration of Deformable Organs in Medical Volume Data by Exhaustive Search

Abstract

This paper proposes a novel framework for fully automatic localization of deformable organs in medical volume data, which can obtain not only the position but also simultaneously the orientation and deformation of the organ to be searched, without the need to segment the organ first. The problem is defined as one of minimizing the sum of squared distances between the organ model’s surface points and their closest surface points extracted from the input volume data. The geometric alignment, or so-called registration, of three-dimensional models by least square minimization always has the problem of initial states. We argue that the only way to solve this problem is by the exhaustive search. However, the exhaustive search takes much computational cost. In order to reduce the computational cost, we make efforts in the following three ways: (1) a uniform sampling over 3D rotation group; (2) Pyramidal search for all parameters; (3) Construction of a distance function for efficiently finding closest points. We have finished experiments for searching the six parameters for position and orientation, and the results show that the proposed framework can achieve correct localization of organs in the input data even with very large amounts of noise. We are currently expanding the system to localize organs with large deformation by adding and searching parameters representing scaling and deformation.