A deep learning approach for predicting linear accelerator output settings in automated radiotherapy planning of
Mathieu Gaudreault1,2, Lachlan McIntosh1, Katrina Woodford1,2,3
1Peter MacCallum Cancer Centre, Melbourne, Victoria 3000, Australia.
Background And Purpose:
The monitor units (MU) per control point (CP) control the necessary fine-tuned ablative dose for hypofractionated radiotherapy of oligometastatic cancer. We aimed to introduce strategies maximising the sample size to accurately predict the MU per CP with artificial intelligence (AI).
Materials And Methods:
The 40/68/88 treatment plans of consecutive patients treated between 01/2019 and 06/2024 at our institution for metastatic cancer to the liver/bone/lung were included. Two approaches were considered to maximise the sample size. In one approach, the samples of each anatomical site were extensively augmented to predict the MU per CP from the dose distribution per CP, providing the MU per beam and meterset weight per CP. In the other approach, all samples from all anatomical sites were combined for training. The number of achieved clinical goals based on dose-volume calculation metrics in AI radiotherapy plans (AI-RTPlan) was compared with the number of achieved clinical goals in the clinical plans.
Results:
The mean absolute percentage error between predicted and clinical MU per beam/meterset weight per CP was less than 6.2%. All AI-RTPlans were generated in less than 5 s. At least 90%/5% of patients had the same, or more, achieved clinical goals with AI-RTPlans. Target coverage and dose to organs at risk metrics were within ± 2% and ± 2.3 Gy of the clinical value in all patients, respectively.
Conclusions:
Augmenting data extensively and combining anatomical sites were equivalent and proficient strategies to predict machine settings for radiotherapy planning of oligometastatic cancer.
