DoseBridge AI predicts proton therapy doses with beam-aware diffusion

Radiotherapy dose prediction models typically rely only on CT images and anatomical structures, ignoring beam geometry—a critical factor in intensity-modulated proton therapy (IMPT). A new paper from Zerun Zhang and colleagues at Beaumont Health introduces DoseBridge, a denoising diffusion bridge model that treats the patient CT as a structured bridge endpoint and fuses a spatially aligned beam mask with multiscale representations of CT, target volumes, organs at risk, and beam geometry. This adds just 1.95% parameters over a standard architecture, keeping the model lightweight for small clinical datasets.

DoseBridge was tested on 52 lung cancer patients (42 training, 10 testing) treated with 60 Gy in 30 fractions. It achieved a mean absolute error of 4.170 Gy, a peak signal-to-noise ratio of 23.06 dB, and a structural similarity index of 0.798—outperforming two deep-learning comparison models. Clinical target volume D95 differed from reference by only 0.62 ± 1.6 Gy, and normal-tissue complication probability errors were within 2.2 percentage points for esophagitis and 3.4 for pneumonitis. Notably, changing only the beam mask shifted predicted low-dose entrance regions while preserving high-dose targets, showing the model captures beam-aware planning behavior. The authors position DoseBridge as a feasible planning prior for lung IMPT, pending validation on larger external cohorts.