Exhibit Hall | Forum 3
Purpose: To introduce an AI-based approach that quickly predicts percentile dose values, corresponding to hundreds of uncertainty scenarios, at every voxel allowing for robust evaluation of planning objectives at specific confidence levels for particle therapy.
Methods: We designed a deep learning architecture based on the 3D-UNet that leverages the benefits of dense connectivity and dilated convolutions. Eleven models were trained to predict a set (5th, 10th, 20th, …, 90th, and 95th) of percentile dose distributions from the nominal dose and planning CT. The data consisted of clinical scanning proton plans from 543 prostate cancer patients divided into 332, 104, and 107 plans for training, validation, and testing, respectively. Ground truth percentile values were calculated for each patient from 600 dose distributions representing random setup and range errors. Prediction accuracy was evaluated by computing the (2mm, 2%) gamma passing rate (GPR) and mean absolute difference (MAD) between the predicted and ground truth dose distributions for each test patient. The MAD was also used to evaluate DVH curves for the 5th and 95th percentiles, which we proposed as alternatives to worst-case scenario evaluation.
Results: The mean GPR of the 11 models ranged between 99.1±0.3% and 99.7±0.1% while the MAD fluctuated between 0.18±0.02Gy and 0.40±0.05Gy. The predicted DVH for the 5th and 95th percentiles show comparable high accuracy, with the MAD of the relative volumes for both rectum and bladder never exceeding 0.30% for dose values above 10Gy. The MAD of the CTV dose was less than 0.2Gy for all volumes below 99.9%.
Conclusion: The proposed method produces accurate and fast (~2.5s/percentile) predictions with intuitive statistical interpretation and presents a promising way for objective robustness evaluation. By estimating an ensemble of percentile doses, we can find where planning objectives are likely to fail and quantify the corresponding probability of failure.
Not Applicable / None Entered.