Abstract
PURPOSE/OBJECTIVE(S)
Stereotactic ablative radiotherapy (SABR) has recently been used to treat increasing numbers of lung metastases, either synchronously or over multiple courses of treatment. Selecting the optimal dose and fractionation to balance risk of local failure and treatment toxicity is challenging. This project uses machine learning to provide rapid dosimetry predictions of SABR to multiple lung lesions, allowing exploration of different dose prescription options prior to the radiation (RT) planning process.
MATERIALS/METHODS
A generative adversarial network (GAN) was trained to predict the dosimetry of multi-lesion thoracic SABR treatment from the planning CT scan, target and organ at risk contours, and the prescribed dose-fractionation without the need to carry out treatment planning. RT plans of patients who received at least one SABR treatment for ≥2 lung lesions between 2014-2020 at a single tertiary center were included in the analysis. All prescriptions were converted to their equivalent doses in two Gray fractions (EQD2). SABR treatments received at different timepoints were registered, and EQD2 doses were accumulated with no repair. Model performance was assessed using 5-fold cross validation. Plans were randomly divided into 5 folds, stratified by the number of lesions (no patients crossed folds). Each fold served as the testing set once, with the model trained on the other 4 folds. The model was evaluated on the difference in the volume of lung receiving above 20 Gray (V20) of the predicted dosimetry compared to the true dosimetry.
RESULTS
Treatment plans (n = 103) were included from 81 patients with 280 lesions (62, 23, 8, and 10 plans had 2, 3, 4, and ≥5 lesions respectively). Fifty-five, 18, and 4 patients had a single, 2 and 3 courses of RT respectively. Fifty-two patients were treated for primary lung cancer, 28 patients treated for metastases from other sites, and 1 patient for both. Seven patients (8.6%) developed ≥ Grade 2 pneumonitis. Doses prescribed were 60/8 (n = 136), 55/5 (n = 49), 54/3 (n = 27), 24/1 (n = 21), 35/5 (n = 13), 30/5 (n = 9), and other (n = 25). The mean lung V20 for all patients was 11.3% [1.5%-29.6%]; the mean lung V20 was 9.3%, 12.2%, 13.5% and 20.4% for plans with 2, 3, 4, ≥ 5 lesions respectively. The mean absolute difference (MAD) in lung V20 between the predicted dosimetry and true dosimetry over all 5-folds was 1.9% [0.0%-13.4%]; the MAD in lung V20 between the predicted and actual dosimetry was 1.6%, 2.1%, 2.9% and 2.6% for plans with 2, 3, 4, ≥ 5 lesions respectively.
CONCLUSION
The GAN-based model created in this study can predict the dosimetry of any number of lesions in the thorax treated with SABR. The model can be used to quickly determine the feasibility of SABR treatment for multiple synchronous lesions, or in the retreatment setting. The ability to explore the dosimetry of different prescription options for a given patient prior to RT planning may allow for personalized risk-adapted treatment if combined with local control and toxicity modelling.
Collapse