Purpose: To validate machine-learning (ML) treatment planning on iterative-reconstructed cone beam computed tomography (CBCT) imaging for the purpose of adaptive radiation therapy (ART) for prostate cancer patients receiving hypofractionated (60 Gy in 20 fractions) therapy.
Methods: We first generated a CBCT-to-density table to generate the electron density values required for dose computation. This was achieved by performing deformable registration between CT and CBCT images collected from 5 prostate cancer patients and employing histogram analysis to determine the correlation between CT and CBCT voxel values. We validated dose calculation accuracy by comparing gamma metrics between CBCT and warped CT images on a validation set of 20 patients. Next, we applied a clinically validated CT-based ML treatment planning method to both the CT and CBCT images of 28 separate patients. All plans were created and optimized automatically. CT and CBCT treatment plans were compared for dose equivalence and overall clinical acceptability using institutional dose-volume protocol metrics.
Results: Dose calculation on CBCT images demonstrated excellent accuracy, with 99.7 ± 0.8% of all irradiated voxels within 1% agreement with CT, averaged over all patients. CT and CBCT treatment plans were equivalent for all dose-volume criteria, except for rectum D30, and D1cc of the bowels and penile bulb (p<0.05), which showed decreased median doses on CBCT plans. There were no significant differences in dose-volume criteria pass rates between CT and CBCT plans (p<0.05). Using the signed distance transform revealed no significant differences in the distances between organs at risk and the clinical target volume, suggesting similar geometries.
Conclusion: We demonstrate equivalence in calculated dose and treatment plan quality between CBCT and CT ML plans, without ML retraining on CBCT. Future work will apply this approach for online adaptive planning to demonstrate the value of daily ART for reducing margins and adapting to changing daily anatomy.
Funding Support, Disclosures, and Conflict of Interest: This study is supported by the Collaborative Health Research Programs Grant, jointly funded by the Canadian Institute of Health Research (CIHR) and Natural Science and Engineering Research Council of Canada (NSERC). Chris McIntosh and Thomas Purdie receive royalties from RaySearch Inc. for machine learning based treatment planning methods.