Deformable Breast MRI and CT Registration for Radiation Treatment Planning Using B-spline With Rigidity and Feature Point Distance Regularization

Abstract

To develop a software tool for accurate and automatic deformable registration between MRI and CT breast images, as particularly preferred in preoperative breast radiation therapy planning. A novel B-Spline registration approach of using feature point distance and local rigidity transformation measure as regularization is adopted to correct fine structure inaccuracy and bone area warping artifact due to B-Spline method's lack of physical property modeling and global intensity based similarity metric. The software tool uses Insight Toolkit (ITK) for image input/output, filtering, feature points identification and auto-segmentation. A B-Spline registration toolkit with invertibility regularization was adopted for mutual information similarity measure and feature point distance penalty. It includes the following steps and functionalities: (1) accepting input of images of different modalities and performing initial rigid registration. (2) Finding feature point pairs, and filtering them by bifurcation identification and distance histogram. (3) Segmenting bone, air, and soft tissue in three levels, in CT by Hounsfield value, or in MR by distinguishing piecewise zero-bound intensity areas' volume and location. Three levels of stiffness are assigned to the segmentation result for rigidity regularization. (4) Using multi-scale approach to update B-Spline coefficients and using conjugate gradient for the solver iteration. (5) Within every step, updating feature point's new position from the deformation field for distance penalty, updating mutual information and transformation Jacobian for regularization. MRI and CT data acquired for three representative breast cancer patients were used to test the software tool. The MRI T1 images were registered upon CT images. Feature point pair distances were evaluated among rigid registration, simple B-Spline registration and our proposed registration. The newly developed deformable multimodality image registration tool can accurately register MRI to CT for breast cancer patients without user interaction. The average feature point pair distance was reduced from 6.7 mm to 3.2 mm by simple B-Spline registration, and was reduced to 2.1 mm by the proposed registration method. The bone area warping artifact in deformed MRI also disappeared in the proposed method. The resulting feature point pair distance is satisfactory considering the resolution of MR images is 1.3 mm. Multimodal images (MRI and CT) for breast cancer patients, which often have large soft tissue deformation, were successfully registered to an acceptable accuracy by using the newly developed software tool, enabling RT planning with MRI-based targeting.