Author: Nrusingh C. Biswal, Matthew J Ferris, Michael J. MacFarlane, Jason K Molitoris, Byong Yong Yi, Mark J. Zakhary 👨🔬
Affiliation: University of Maryland School of Medicine, Department of Radiation Oncology, University of Maryland School of Medicine, University of Maryland 🌍
Purpose: Proton head-and-neck treatment plans often struggle to maintain plan quality over the course of treatment due to tumor response, weight-loss, and setup variability. Plan robustness to these changes might be improved by incorporating synthetic CTs (produced by a generative machine learning model) simulating likely future anatomical changes into the robust optimization problem. However, this may also reduce plan quality considerably. To evaluate whether this workflow has merit, an initial feasibility study was performed assessing the plan quality and robustness produced by this workflow when using the patient’s mid-of-treatment CT (viz. synthetic CTs with zero simulation error).
Methods: 10 previously treated patients were randomly selected who had routine (typically weekly) verification CTs (vCT) and substantial anatomical changes warranting replanning. A new plan was generated by robustly optimizing the plan on both the planning CT and the vCT acquired closest to mid-of-treatment. Dose metrics (CTV: D95, V99, V95; serial OARs: D0.03cc; parallel OARs: DMean) were compared between the clinical and new plan, when calculated on the planning CT (evaluating nominal plan quality) and on each of the vCT scans (evaluating plan robustness).
Results: The workflow had a negligible impact on the nominal plan quality, with all CTV and OAR dose metrics between plans differing by less than 0.3% or 50 cGy, on average. When calculated on the vCT, the new plans sustained higher CTV coverage (V95: +0.4-1.1%, V99: +2.2-4.1%) and exhibited more gradual OAR dose increases – when increases occurred. If treated with the new plan, 3 patients would likely have completed treatment without plan revision, and 4 would likely have plan revisions performed 1-2 weeks later than the clinical plan.
Conclusion: Incorporating the mid-of-treatment CT into the robust optimization impacted nominal plan quality minimally while improving robustness, demonstrating potential for a future workflow that utilizes synthetic CTs predicting anatomical changes.