The purpose of this study was to develop a workflow process that enables quantitative assessment of different image registration techniques used for head and neck simulation CT to diagnostic CT coregistration.
Abdallah S. Mohamed, MD, MSc, Manee-Naad Ruangskul, MD, Musaddiq J. Awan, MD, Charles A. Baron, MD, Richard Castillo, PhD, Edward Castillo, PhD, Thomas M. Guerrero, MD, PhD, Esengul Kocak-Uzel, MD, Jinzhong Yang, PhD, Laurence Court, PhD, G. Brandon Gunn, MD, Adam S. Garden, MD, David I. Rosenthal, MD, Clifton D. Fuller, MD, PhD; UT MD Anderson Cancer Center
Background and Purpose: Developing a framework to validate the performance of image registration algorithms is critical before application for tumor localization and therapeutic targeting. The purpose of this study was to develop a workflow process that enables quantitative assessment of different image registration techniques used for head and neck simulation CT (SimCT) to diagnostic CT (DxCT) coregistration.
Materials and Methods: A total of 68 reference anatomic regions of interest (ROIs) were manually contoured on each of 11 paired SimCTs and DxCTs of head and neck patients treated with definitive intensity-modulated radiotherapy (IMRT). DxCT was registered to SimCT rigidly and through four different deformable image registration (DIR) algorithms: Atlas-based, b-spline, demons, and optical flow. The resultant deformed ROIs were compared with manually contoured reference ROIs using similarity coefficient metrics (ie, Dice similarity coefficient) and surface distance metrics (ie, 95% maximum Hausdorff distance).
Results: All DIR algorithms showed improved performance over rigid registration for all used comparison metrics (Steel test: P < .008 after Bonferroni correction), excepting optical flow for surface distance metrics. The Atlas-based algorithm had the best DIR performance (mean Dice of 0.65 ± 0.15, mean false-negative Dice of 0.11 ± 0.18, mean false-positive Dice of 0.58 ± 0.26, and mean 95% maximum Hausdorff distance of 6.79 mm ± 7.6). The performance of different algorithms varied substantially for specific anatomic ROIs and subgroups. Overall, the performance of most algorithms was better in matching bony and cartilaginous ROIs than muscular, glandular, vascular, and other soft tissue ROIs.
Conclusions: Development of a formal ROI-based quality assurance workflow for registration assessment revealed improved performance with DIR techniques over rigid fusion and provided head and neck ROI-specific benchmarks for DxCT-SimCT coregistration for future efforts. After QA, DIR implementation should be the standard for head and neck DxCT-SimCT allineation.