PyDoseRT Photon Adds Small Neural-Prior Corrections to a Differentiable Physics Engine for Radiotherapy Dose Calculation

Researchers at Umeå University and the Medical University of Vienna have introduced PyDoseRT Photon, an attempt to strike a balance between a faster but more approximate pencil-beam radiation dose engine and a neural corrector that produces results closer to Monte Carlo calculations. The system implements a GPU-based differentiable physics engine in PyTorch. It first calculates a baseline dose using beam quality, radiological depth, and TERMA scaling, then applies data-driven corrections layer by layer instead of using a large black-box network to predict the complete dose field directly from images.

The first neural prior is a two-dimensional fluence corrector with approximately 39,000 parameters. The second has only 48 parameters and corrects lateral tissue heterogeneity using a mass-preserving Gaussian redistribution operator. Both are trained through the frozen physics engine, and their parameters are fixed after training. Finally, a 1.36-million-parameter 3D U-Net predicts a bounded multiplicative gain and an additive residual from seven channels for each control point in beam’s-eye-view coordinates. Every learning stage uses zero initialization, making its starting point equivalent to the uncorrected analytical solution and reducing the risk that the model will initially disrupt the physics baseline.

For the MRI pathway, nnU-Net first synthesizes a CT image, after which the same dose engine and correctors are used. On local CT and MR validation sets, respectively, the paper reports plan MAEs of 0.0024 and 0.0047, along with 1%/1 mm gamma pass rates of 99.12% and 96.95%. Full-case computation took approximately 46 and 49 seconds. These figures come from a submission configuration designed for the DoseRAD2026 challenge, whose ranking metric places relatively high weight on speed, and therefore cannot be interpreted directly as showing that the method is superior to Monte Carlo approaches for clinical treatment planning.

The foundational PyDoseRT engine and test programs have been released publicly, but the research summary does not establish that the Photon model weights, training data, and complete reproduction workflow have all been released. The next steps should include external validation across machine types, institutions, and cases with extreme tissue heterogeneity, as well as separate measurements of how the analytical core, neural priors, and U-Net each contribute to error and latency.