Comparative evaluation of regression methods for 3D-2D image registration

Abstract

We perform a comparative evaluation of different regression techniques for 3D-2D registration-by-regression. In registration-by-regression, image registration is treated as a nonlinear regression problem that relates image features of 2D projection images to the transformation parameters of the 3D image. In this work, we evaluate seven regression methods: Multiple Linear and Polynomial Regression (LR and PR), k-Nearest Neighbour (k-NN), Multiple Layer Perceptron with conjugate gradient optimization (MLP-CG) and with Levenberg-Marquardt optimization (MLP-LM), Radial Basis Function network (RBF) and Support Vector Regression (SVR). The experiments are performed using simulated X-ray images (DRRs) of nine coronary vessel trees, allowing us to compute the mean target registration error (mTRE) to the ground truth. All methods were robust to large initial misalignment and the highest accuracy was achieved using MLP-LM and RBF. © 2012 Springer-Verlag.