2071 - Machine Output Settings Prediction for Volumetric Modulated Arc Therapy Artificial Intelligence Treatment Planning of Small Tumors in the Lung
Presenter(s)
M. Gaudreault1,2, V. Panettieri1,2, K. Woodford2, J. Li2, S. Harden1,2, S. Porceddu1,2, and N. Hardcastle1,2; 1Sir Peter MacCallum Department of Oncology, The University of Melbourne, Melbourne, VIC, Australia, 2Peter MacCallum Cancer Centre, Melbourne, VIC, Australia
Purpose/Objective(s): Volumetric modulated arc therapy (VMAT) artificial intelligence (AI) treatment planning aims to generate deliverable treatment plans without dose optimization. In this concept, linear accelerator settings are predicted from minimal input by sequential or integrated AI networks, to reduce time to treatment and improve plan quality. The monitor units (MU) per control point (CP) drive the dose intensity and their accurate predictions are essential in VMAT AI treatment planning. We hypothesize that AI can rapidly and accurately predict the MU per CP for treatment planning of small lung tumors.
Materials/Methods: Consecutive patients treated for either primary lung cancer or lung metastases with either 28 Gy in one fraction or 48 Gy in four fractions in a single institution between 01/2019 and 11/2024 were considered. We constructed a deep learning network with an encoder-bottleneck architecture. The network inputs were two-dimensional average dose intensity projection (ADIP) images, generated by averaging three-dimensional dose distribution per control point (CP) along the beam-eye view direction. The predicted MU per CP were converted to MU per beam and meterset weight per CP to create an AI-radiotherapy plan (AI-RTPlan), in which all other clinical machine settings were retained. The AI-RTPlans were imported into the treatment planning system (TPS) for dose calculation. The calculated dose distributions from AI-RTPlans were compared with the clinical dose distributions with gamma passing rates (?PR) at 3% and 2 mm, clinically achieved goals, and dose-volume metrics. The relative difference of conformity indices CI100 and CI50 between AI-RTPlans and clinical plans was reported.
Results: In 208 patients, 34,224 / 8,556 / 10,120 ADIPs and their associated MU per CP from 164 / 41 / 50 clinically delivered plans were used in the training / validation / testing dataset. The absolute percentage error between predicted and clinical MU per beam and meterset weight per CP was lower than 4.4% and 5.4%, respectively. Each AI-RTPlan was generated in less than three seconds. Between AI-RTPlans and clinical plans, the median (min,max) ?PR(3%,2mm) was 100% (89%,100%). All 50 plans in the testing dataset had a greater or equal number of achieved goals with AI-RTPlans. Improvements in target coverage were observed in three patients. The absolute differences in organs at risk dose between AI-RTPlans and clinical plans were within ±3 Gy. The conformity indices CI100 and CI50 were 0.9% and 0.5% larger (p-values < 0.001 in both cases) in AI-RTPlans than in clinical plans, respectively.
Conclusion: The MU per CP can be rapidly predicted with high accuracy, resulting in equivalent or improved plan quality. The deep learning methods introduced in this work can be integrated into a larger workflow for automated VMAT AI treatment planning of small tumors in the lung.