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

Background Accurate 3D dose calculation for Pencil Beam Scanning Proton Therapy (PBSPT) is typically performed using Monte Carlo (MC) dose engines. Although MC remains the gold standard for dosimetric accuracy, its high computational cost limits applicability in adaptive workflows and other tasks requiring repeated recalculation. Deep learning methods have recently emerged as a promising alternative for fast and accurate proton dose computation. Purpose and Methods We address a major limitation of existing deep learning proton dose engines: the requirement that each beamlet be sampled in its beam’s-eye-view, with the proton beam direction orthogonal to the lateral image planes, which requires a local beam’s-eye-view resampling of the computed tomography (CT) patch for every beamlet. To overcome this, we propose the Angle-dependent Dose Transformer Algorithm (ADoTA), which augments the model with an additional input channel encoding a fast analytical beamlet shape projection. The projection supplies explicit beam direction information, so that a single field-level CT transformation remains: the CT is rotated and resampled once per irradiation field to an isotropic 2 × 2 × 2 mm 3 grid, and each beamlet is then obtained by cropping the relevant sub-volume from this shared grid. The model was trained on CT data from 108 patients to predict beamlet dose distributions for initial energies from 70 to 150 MeV (thoracic) and 70 to 270 MeV (abdominal/pelvic) over an 80 × 110 mm 2 field and evaluated on an independent, non-overlapping test cohort of 50 patients. Results On the independent test set, ADoTA achieved gamma pass rates of 99.40 ± 0.86% and 99.87 ± 0.23% for thoracic and abdominal/pelvic cases, respectively, under the Γ(1%, 3 mm, 0.1%) criterion. Batched inference achieved an effective throughput equivalent to 3.14 ± 0.31 ms per beamlet at a batch size of 200, whereas a beamlet processed on its own required 14.1 ms. By resampling each field once to a 2 × 2 × 2 mm 3 grid and then only cropping the relevant sub-volume per beamlet, ADoTA avoids the beamlet-specific local beam’s-eye-view resampling required by prior methods. In a 5640-spot prostate-plan benchmark with both routes optimized on the graphics processing unit (GPU), the analytical beamlet shape projection was 3–4 × cheaper on compute and approximately 6–7 × cheaper including host-device transfer than batched local beam’s-eye-view sampling, per beamlet, reducing the plan’s direction-encoding cost by 9.6 × and its total compute time by 1.48 × . For full treatment plans Γ(2%, 2 mm, 10%) pass rates averaged 95.1% for lung plans and 96.1% for prostate plans, while the target near-minimum dose D 98% was systematically underestimated by 1.8–2.5% of prescription in prostate plans and by 4.2–5.9% in thoracic plans. Conclusions ADoTA is a beamlet angle-aware deep learning proton dose engine that avoids beamlet-specific local beam’s-eye-view resampling, retaining a single field-level CT transformation. It achieves close to Monte Carlo-level accuracy at the beamlet level and high accuracy at the full-plan level, with a clinically relevant target underdosage in thoracic plans.