Author: Jeremy S. Bredfeldt, Benito De Celis Alonso, Braian Adair Maldonado Luna, Kevin M. Moerman, Gerardo Uriel Perez Rojas, René Eduardo Rodríguez-Pérez, Kamal Singhrao 👨🔬
Affiliation: Department of Radiation Oncology, Brigham and Women's Hospital, Harvard Medical School, Department of Radiation Oncology, Brigham and Women's Hospital, Dana Farber Cancer Institute, Harvard Medical School, Department of Mechanical Engineering, University of Galway, Faculty of Physics and Mathematics, Benemérita Universidad Autónoma de Puebla 🌍
Purpose: Online adaptive radiotherapy replanning for single-isocenter bone cancer metastasis treatment reduces on-table treatment time and patient discomfort compared to the multi-isocenter standard-of-care. The current standard-of-care uses deformable image registration (DIR) to adapt radiotherapy contours to on-table anatomy, often causing unphysical deformations requiring manual corrections. This study develops a novel modality-independent biomechanical (BM) model based on finite element analysis using FEbio and compares its performance in pelvic bone structures (PBS) registration against rigid registration (RR) and Demons method combined with RR (D+RR).
Methods: A male pelvis dataset with 16 patients was employed, each with a fixed planning CT image and a target image based on a treatment-day MRI-derived synthetic CT. These model´s accuracy was quantified using Dice coefficient and mean-absolute-error. For RR, PBS were segmented from the original images, then RR was applied. For D+RR, the original DICOM images were registered via RR and Demons. PBS were segmented for comparison.
For BM+RR, RR was applied to PBS segmentation and processed individually to generate each bone´s mesh within the same z-axis range. An iterative process determined the transformations to align both states, nodal displacements for the mesh surface were calculated to improve registration. A finite element simulation using GIBBON toolbox, applied these transformations, generating each bone´s deformation vector field (DVF). The DVFs were combined and applied to a 3D image of the initial PBS, obtaining a warped image in the target position.
Results: The study found Dice values of 0.87±0.03, 0.82±0.02 and 0.78±0.17 for BM+RR, RR and D+RR respectively. For mean-absolute-error the values were 4.0e-3±1e-3, 5.0e-3±1e-3 and 1.7e-2±1e-2 respectively.
Conclusion: This study demonstrated that the BM model outperformed both RR and DIR-based methods for the creation of DVFs based on multimodal image sets. Future work will assess the dosimetric accuracy of single-isocenter plans deformed using the BM model.