BeamGAN: a physics-informed conditional generative adversarial network for Monte-Carlo-free portal dose image prediction from incident fluence maps

Abstract Objective. To develop and validate a Monte-Carlo-free deep learning framework that predicts EPID portal dose images directly from treatment-planning-system (TPS) exported incident fluence maps, enabling real-time patient-specific quality assurance (QA) without Monte Carlo (MC) simulation or empirical per-machine commissioning. Approach. We propose BeamGAN, a physics-informed conditional generative adversarial network. A three-channel input—TPS incident fluence, its Sobel gradient map, and a beam's-eye-view (BEV) target-thickness projection ray-traced through the planning CT with the PTV mask—embeds geometric and edge-scatter priors into an eight-level U-Net generator with a PatchGAN discriminator and an auxiliary registration network. The framework was trained on 91 head-and-neck IMRT patients (split 64/15/12) from one Elekta Versa HD linac, benchmarked against a Brownian-Bridge diffusion baseline (BiBBDM), and evaluated by input-channel ablation and error-injection sensitivity tests. Main results. BeamGAN achieved gamma passing rates (GPR) of 97.40 ± 0.67% (3%/3 mm), 95.42 ± 0.99% (3%/2 mm) and 87.00 ± 1.63% (2%/2 mm; 10% threshold), outperforming BiBBDM (83.32 ± 2.43%, 77.08 ± 3.20% and 67.20 ± 2.45%). On the held-out test set (n = 12), the median GPR was 95.47% (IQR 94.86–96.27) at the TG-218 primary criterion 3%/2 mm; 8/12 cases met the institutional ≥95% acceptance threshold and all 12 exceeded the 90% action level. Ablation identified the fluence-gradient channel as the dominant contributor, with the anatomical prior providing complementary gains. Sensitivity analysis established a criterion-dependent error hierarchy: systematic MLC errors were most detectable (ΔGPR up to 21.23% at 1%/1 mm), followed by collimator rotation (10.31%), whereas MU scaling and setup errors produced marginal effects. BEV computation required 0.3–0.5 s per beam and inference under 50 ms. Significance. Embedded geometric priors can approximate explicit MC scatter modelling in portal dose prediction, removing a principal computational barrier to real-time EPID-based verification within the validated configuraiton. Multi-institutional validation is warranted before clinical deployment.