Room: Exhibit Hall | Forum 6
Purpose: To propose an automatic and patient-specific approach to evaluate the spatial accuracy of multi-modality deformable image registration (DIR), and use it to evaluate DIR accuracy on Halcyon's megavoltage cone-beam CT (MVCBCT).
Methods: 396 MVCBCT images and 15 CT images from 15 historical patients treated on Halcyon were retrospectively selected. Simulated deformation vector field (Vf0) was introduced to warp the kVCT and MVCBCT images. Using B-spline DIR algorithm, warped (original) CT images were registered to all original (warped) MVCBCT images of the same patient, where the registration and resulted transformation vector fields were denoted as DIR1 (DIR2) and Vf1 (Vf2). Ground truth value for composite transform DIR2 - DIR1, where the minus sign referred to adding the inverse transform, was then mathematically equal to doubled Vf0, and the registration error was quantified as the mean vector length of (Vf2 - Vf1 - 2*Vf0). Mean vector length of (Vf2 - Vf0) was also computed for comparison. The whole procedure was automated by open-source software plastimatch.
Results: The registration error (mean vector length of Vf2 - Vf1 â€“ 2*Vf0) averaged over fractions for each patient ranged from 0.089mm to 1.627mm. Despite the inferior image quality of MVCBCT, the mean registration error was 0.538mm and over 80% of fractions had registration errors within 1mm. Assuming patient anatomy as rigid body and neglecting intrinsic differences between imaging modalities, the â€œground truthâ€? deformation would be Vf0 and the mean registration error (mean vector length of Vf2 â€“ Vf0) would be mistakenly enlarged to 6.516mm.
Conclusion: Traditional virtual deformation method could not be directly applied to multi-modality DIR evaluation due to the existence of unknown intrinsic deformation of patient and between imaging modalities. By simple extra processing, our method could provide a more objective evaluation of multi-modality DIR spatial accuracy on a patient-specific basis.
Funding Support, Disclosures, and Conflict of Interest: This work was jointly supported by Capital's Funds for Health Improvement and Research (2018-4-1027), Beijing Natural Science Foundation (7172048, 1174016 and 1184014).