KEY POINTS
- The Angle-dependent Dose Transformer Algorithm (ADoTA) is a deep-learning proton dose engine designed to predict individual pencil-beam dose distributions without separate beam’s-eye-view CT resampling for every spot. It was developed using CT data from 108 patients, independently tested on 3,500 beamlets from 50 patients, and further evaluated using lung and prostate treatment plans.
- ADoTA combines CT information, proton energy, and an analytical Gaussian projection encoding beamlet orientation. CT is transformed once per irradiation field into a 2 × 2 × 2 mm³ grid, avoiding repeated beamlet-specific geometric transformations.
- Beamlet-level gamma passing rates under 1%/3 mm criteria were 99.40 ± 0.86% for thoracic and 99.87 ± 0.23% for abdominal/pelvic anatomy. At 1%/2 mm, passing rates were 97.86% and 99.46%, respectively, indicating greater sensitivity to heterogeneous thoracic tissue.
- Mean absolute distal R80 error was 1.66 mm in thoracic versus 0.57 mm in abdominal/pelvic beamlets. Absolute range errors exceeded 3 mm in 10.0% of thoracic beamlets and 5 mm in 5.4%, compared with no such exceedances in the abdominal/pelvic cohort.
- GPU inference achieved 3.14 ± 0.31 ms per beamlet with batch size 200. In a 5,640-spot benchmark, analytical direction encoding reduced geometric preparation cost by 9.6-fold and total computation time by 1.48-fold compared with optimized beamlet-specific resampling.
- For full treatment plans, mean 2%/2 mm gamma passing rates were 95.1% in lung and 96.1% in prostate cases. Nevertheless, target D98 was systematically underestimated by 4.2–5.9% of prescription in thoracic plans and 1.8–2.5% in prostate plans.
- The largest evaluated plan contained 9,273 spots and was reconstructed in 52.11 seconds. Generalizability to other proton beam models and clinical systems remains uncertain, and the systematic target-dose underestimation is a major obstacle to replacing Monte Carlo dose calculations.
CLINICAL TAKEAWAY
ADoTA demonstrates that angle-aware deep learning can accelerate proton dose calculation by eliminating redundant beamlet-specific CT transformations, potentially benefiting adaptive and robust optimization workflows. However, systematic underestimation of target coverage, particularly in heterogeneous thoracic anatomy, remains a clinically important limitation despite high gamma passing rates, and further validation is needed before clinical deployment.