Angle-Dependent Dose Transformer Algorithm for Fast Proton Therapy Dose Calculations

Abstract

Monte Carlo (MC) dose engines are the gold standard for pencil beam scanning proton therapy, but their computational cost limits adaptive workflows and other tasks requiring repeated dose recalculation. Existing deep learning beamlet dose engines require each beamlet to be sampled in its beam's-eye-view, which implies a local resampling of the computed tomography (CT) patch for every beamlet. We propose the Angle-dependent Dose Transformer Algorithm (ADoTA), which adds an input channel encoding a fast analytical beamlet shape projection that supplies the beam direction explicitly. The CT is resampled only once per irradiation field to an isotropic $2\times 2\times 2$ mm${}^3$ grid, and each beamlet is obtained by cropping this shared grid. The model was trained on CT data from 108 patients for initial energies of 70 to 150 MeV (thoracic) and 70 to 270 MeV (abdominal/pelvic) and evaluated on an independent test cohort of 50 patients. ADoTA achieved $\Gamma (1\%,3\mathrm{mm},0.1\%)$ pass rates of $99.40\pm 0.86\%$ (thoracic) and $99.87\pm 0.23\%$ (abdominal/pelvic), with an effective batched throughput of $3.14\pm 0.31$ ms per beamlet. In a 5640-spot prostate plan benchmark with both routes optimized on the GPU, the analytical projection reduced the plan's direction-encoding cost by $9.6\times$ and its total compute time by $1.48\times$ compared with local beam's-eye-view sampling. For full treatment plans, $\Gamma (2\%,2\mathrm{mm},10\%)$ pass rates averaged 95.1% (lung) and 96.1% (prostate), while the target $D_{98\%}$ was underestimated by 1.8 to 2.5% of prescription in prostate plans and 4.2 to 5.9% in thoracic plans. ADoTA achieves close to MC accuracy at the beamlet level without beamlet-specific resampling, with a clinically relevant target underdosage in thoracic plans that should be resolved before clinical use.

arXiv commenthttps://github.com/Medical-Physics-Technology/adota