Mixture of Principal Axes Registration for Change Analysis in Computer-Aided Diagnosis

R Srikanchana,K Huang,JH Xuan,M Freedman,Y Wang

DOI: https://doi.org/10.1109/aipr.2001.991199

2001-01-01

Abstract:Non-rigid image registration is a prerequisite for manymedical image analysis applications such as image fusionof multi-modality images and quantitative change analysisof a temporal sequence in computer-aided diagnosis. By establishingthe point correspondence of the extracted featurepoints, it is possible to recover the deformation using non-linearinterpolation methods such as the thin-plate-splineapproach. However, it is a difficulty task to establish exactpoint correspondence due to the high complexity of nonlineardeformation existed in medical images. In this paper,a mixture of principal axes registration (mPAR) method isproposed to resolve the correspondence problem through aneural computational approach. The novel feature of mPARis to align two point sets without needing to establish explicitpoint correspondence. Instead, it aligns the two pointsets by minimizing the relative entropy between their probabilitydistributions resulting in a maximum likelihood estimateof the transformation matrix. The registration processconsists of: (1) a finite mixture scheme to establish animproved point correspondence and (2) a multilayer perceptronneural network (MLP) to recover the nonlinear deformation.The neural computation for registration used acommittee machine to obtain a mixture of piece-wise rigidregistrations, which gives a reliable point correspondenceusing multiple extracted objects in a finite mixture scheme.Then the MLP is used to determine the coefficients of a polynomialtransform using extracted cross points of elongatedstructures as control points. We have applied our mPARmethod to a temporal sequence of mammograms of a singlepatient. The experimental results show that mPAR not onlyimproves the accuracy of the point correspondence but alsoresults in a desirable error-resilience property for controlpoint selection errors.