Purpose: Recently a deep-learning model was proposed to boost the accuracy of AAA doses to the level of AcurosXB doses (AXB). A robustness study was performed to evaluate the model on plans across different sites, planning techniques, and machine models to examine its generalizability.
Methods: In total 66 AAA plans of 5 sites from the same institution were recalculated using the AXB dose engine, and the resulting AAA-AXB dose pairs were used as the testing cohort. A prior deep-learning model trained and validated using VMAT lung plans on Elekta LINACs (VersaHD) was deployed on the testing cohort to derive deep-learning-boosted doses (DLB-AAA) from original AAA doses. The DLB-AAA doses were evaluated using gamma passing rates with the calculated AXB doses as reference. The testing cases were grouped into VMAT-only, same machine model-only, and VMAT & same machine model-only to evaluate the effects of planning techniques and machine models on the deep-learning model accuracy.
Results: When comparing the gamma passing rates for all testing cases, the DLB-AAA doses do not show better match to AXB doses than the original AAA doses, indicating overall limited model generalizability. However, when restricting the evaluation plans to VMAT-only, the DLB-AAA doses are significantly closer to AXB doses than the original AAA doses across all sites except for head-and-neck (HN), potentially due to the substantial difference in modulation factors between HN and lung VMAT plans. To improve the generalizability of the model, we propose to include plans from all sites and planning techniques for training, with potential fine-tuning through transfer learning warranted.
Conclusion: Our DL-based AAA-AXB dose conversion model shows robustness towards machine models and treatment sites (except for HN), while lacks robustness towards different planning techniques (VMAT vs. IMRT vs. 3D). Transfer learning or a more generalized model is needed to further improve its robustness.
Funding Support, Disclosures, and Conflict of Interest: The study was partially supported by funding from the National Institutes of Health (R01CA240808) and from the University of Texas Southwestern Medical Center
Not Applicable / None Entered.