Feature Line Parallelization for Enhanced Registration of Post-implant CT and MRI Prostate Image Volumes

Abstract

The utility of feature line parallelization as the initial step in manual or semi-automatic 3D rigid-body registration of post-implant CT and MR prostate image volumes is investigated. Image data consisted of CT and 1.5T MRI (balanced fast-field echo, Te = 4.8 ms, Tr = 9.6 ms) axial images of the pelvis for 3 patients acquired approximately 4 weeks after implantation with I-125 seeds. Axial slices were 3 mm thick and spaced 3 mm apart (no gap), with a pixel pitch of ∼0.3 mm for both imaging modalities. Registration was performed in two stages. The first involved obtaining a transformation to parallelize straight lines fit to corresponding features running primarily in the sup-inf (Z) direction in the image volumes. Features selected for analysis were seed trains for CT and needle tracks/seed voids for MR. The second stage consisted of using a combination of the Procrustes algorithm and manual registration, or a normalized mutual information (NMI) algorithm, to obtain the remaining relative X, Y and Z translations and Z-axis rotation required to complete registration. Image registration performed using the Procrustes algorithm and manual approach alone when there are 6 degrees of freedom (DOF) was found to be challenging and time consuming, while 6 DOF NMI registration proved not to be consistently reliable. Feature line parallelization facilitates 3D rigid-body registration by reducing it to a 4 DOF problem involving a single rotation in the information-rich axial planes, and thereby enables successful completion by either manual or semi-automatic methods, as evidenced by comparison of target registration errors. As the initial step in CT and MRI prostate image volume registration for post-implant quality assessment, feature line parallelization affords opportunity to improve both the accuracy and speed of manual and semi-automatic image registration methods.