Purpose: The treatment planning process is time-consuming due to the complexity of iterative optimizations, this bottleneck can be resolved through deep learning dose prediction. However, this process excludes the mechanical parameters necessary for a deliverable plan. We aim to generate a streamlined VMAT planning process for prostate cancer using the Eclipse photon optimizer (PO) to generate a deliverable plan from an automatically predicted 3-dimensional (3D) radiation dose distribution.
Methods: A conditional generative adversarial network, trained with 127 patients, was used to predict normalized 3D radiation dose distributions for 5 prostate cancer patients treated to 79.2Gy in 44 fractions, which were then imported into the treatment planning system. The PO was used to define the optimization parameters based off the predicted dose volume histogram. One optimization was performed using three consistent optimization objectives at predefined points for each OAR (rectum, bladder, penile bulb and right/left femoral heads) and PTV. Key dosimetric parameters were compared between the original planned, predicted and deliverable doses. 3D gamma analysis using plan delivery logs was performed using 3%3mm criteria.
Results: Dosimetric comparison demonstrated that the deliverable doses had improved PTV coverage and comparable dose to OARs to the predicted and planned dose. The PTV coverage (D95%) on average increased from 96.0% to 100.6% between predicted and deliverable dose. The D50%(Gy) and V20Gy(%) for the bladder and rectum on average increased from 27.6Gy and 57.0% to 28.6Gy and 57.4% and 37.3Gy and 78.7% and 39.4Gy and 79.5% respectively. The right and left femoral D5%(Gy) decreased from 29.9Gy to 29.6Gy and 34.0Gy and 33.9Gy, respectively. Mean dose to the penile bulb increased from 26.2Gy to 29.1Gy. The average passing rate for 3D gamma analysis was 99.6%.
Conclusion: With deep learning predicted doses, we can generate a deliverable plan with radiation dose distributions comparable to the predicted dose.
Funding Support, Disclosures, and Conflict of Interest: The work was supported by a Research Scholar Grant, RSG-15-137-01-CCE from the American Cancer Society.
Treatment Planning, Quality Assurance, Inverse Planning
TH- External Beam- Photons: Treatment planning using machine learning/Knowledge Based Planning/automation