Ibón García Gómez

Neural Fast Monte Carlo for PBS Proton Therapy

A differentiable pencil-beam engine with a zero-init residual corrector, on the DoseRAD2026 open benchmark (~81,000 Geant4 beamlets, 85 energies).

Monte Carlo is the reference dose calculation in PBS proton therapy, but too slow for a planning loop. This project builds a differentiable pencil-beam engine (85-energy LUTs, water-equivalent-depth ray-tracing, two-stage MC calibration) and an xLSTM residual corrector with zero-init heads — at step zero its contribution is exactly the physics, so the neural network can only improve on it, never break it.

On 12 held-out patients (277 beamlets, three terrain types), the analytical engine alone reaches 93.3 % (water), 94.7 % (moderate) and 89.4 % (hard) median γ 1%/1mm pass rate at the 10% cut, with ~0.3 mm R80 error — an honest FAIL against the formal 98.8 % gate, with the engine's nuclear-tail limitations documented and deferred to the corrector. Development is test-driven with fail-closed provenance; the results-first repository documents the method and current numbers.

Beam-eye-view dose: Monte Carlo reference vs C1 engine, with difference map and Bragg curve (γ 1%/1mm: 98.6 % @10%, 95.4 % @0.5% — heterogeneous terrain)

Geant4 / OpenGATE MC versus clinical TPS. γ_{3 %/3 mm} = 76.3 %

EBRT Dose Prediction. Monte Carlo Dataset and Differentiable Neural Surrogate

Independent research project, with anonymised clinical data kindly shared by a medical physicist at Hospital Universitario de León. Reached the finalist stage of the UCEM innovation call (2026).

End-to-end pipeline for prostate VMAT radiotherapy. A Geant4 / OpenGATE 10 Monte Carlo simulator generates a physics-validated dataset of single-beam dose projections (9 patients, 179 gantry angles, around three weeks of HPC time on Altamira). A 3D TransUNet + FiLM neural surrogate, with a 4-channel physics-informed input (RED, WED, DTS, TERMA), learns to predict 3D dose in about one second, replacing minutes-to-hours of clinical computation.

Training uses DDP across 2 × NVIDIA A30, AMP-FP16, and a physics-informed composite loss (asymmetric dose-excess, gradient-match, mean-bias). The roadmap extends to a HeteroGNN amortized inverse-planner trained end-to-end through the differentiable surrogate, so that the optimiser converges to the mathematical optimum of the clinical objective without using clinician-approved plans as targets.