An Evaluation of Deformable Image Registration Algorithms for MRI-to-CT for Gynecologic Brachytherapy Applications
Abstract
Purpose
Accurate deformable image registration (DIR) between magnetic resonance imaging (MRI) and computed tomography (CT) remains a significant challenge in gynecologic brachytherapy, particularly between images acquired with and without a brachytherapy applicator in place. This study evaluated the performance of commercially available intensity- and contour-guided algorithms and in-house finite element method (FEM)–based algorithms for MRI-to-CT DIR.
Methods
Six DIR algorithms were evaluated, including three intensity-based methods (Demons, B-Spline, and multi-modality (MM)), one contour-guided (CG) approach, and two FEM-based methods (hybrid (HFEM) and contour-based (CFEM)). MRI images acquired without an applicator were registered to CT simulation images acquired with an applicator in place. Registration accuracy was assessed using contour-based analysis of three organs-at-risk: bladder, rectum, and bowel-sigmoid. Quantitative evaluation employed Dice similarity coefficient (DSC), mean distance to agreement (MDA), and Hausdorff distance (HD).
Results
Across the OARs evaluated, the CG DIR algorithm yielded an average DSC of 0.77 ± 0.03, MDA of 5.5 ± 0.6 mm, and HD of 20.7 ± 5.2 mm. Intensity-based methods (Demons, B-Spline, and multi-modality) produced average DSC values ranging from 0.5-0.6, MDA values ranging from 5.4-5.6 mm, and HD values ranging from 28-35 mm. The CFEM algorithm resulted in an average DSC of 0.60 ± 0.12, MDA of 5.4 ± 0.7 mm, and HD of 56.7 ± 35.0 mm. The HFEM algorithm had an average DSC of 0.27 ± 0.08, MDA of 5.6 ± 1.5 mm, and HD of 63 ± 40 mm.
Conclusion
CG and CFEM-based DIR approaches outperformed alternative methods for MRI-to-CT registration with and without applicators present in gynecologic brachytherapy. Accurate DIR is essential for target delineation and future work toward reliable dose accumulation for treatment evaluation. Hybrid contour- and FEM-based methods show promise for improving registration robustness in complex brachytherapy imaging scenarios involving multi-modal images and large-magnitude deformation.