pyRadPlan-dosecalc-Bayes-proton-lung

Bayesian proton dose predictor (ConvBayes_new + blitz BayesianLSTM). Given a CT cuboid and a bixel energy, it predicts a local dose cuboid together with a Monte-Carlo predictive variance, estimated from an ensemble of seeded forward passes (the BayesianLSTM samples fresh weights from its posterior on every call).

Loadable via pyRadPlan.ml.load_model and usable directly as the AIBeamletEngine dose-calculation engine in pyRadPlan:

pln.prop_dose_calc = {"engine": "AIBeamlet", "model": "pyRadPlan-dosecalc-Bayes-proton-lung"}

Repository contents

Follows the pyRadPlan ML model contract (pyRadPlan.ml):

File	Purpose
model.py	ConvBayes_new network definition
preprocessor.py	ConvBayesEnsemblePreprocessor — input assembly + Monte-Carlo ensemble inference + output scaling
weights.safetensors	Trained weights
model_config.json	Declarative model/preprocessing/dose-calc configuration

Outputs

• physical_dose — Monte-Carlo mean dose
• variance — predictive variance (Gy²) across the ensemble

The ensemble size defaults to 100 (model_preprocessing.ensemble_size in model_config.json), and can be overridden per plan via pln.prop_dose_calc["ensemble_size"].

Training assumptions

• Radiation mode: protons
• Energy range: 75–160 MeV
• Trained machine: Generic
• Sampling grid: 2 mm spacing, 52 mm lateral range, depth range [-220, 100] mm

Predictions outside this range are not guaranteed to be accurate; AIBeamletEngine warns when the plan falls outside the declared range.

Security note: loading this model executes model.py/preprocessor.py shipped in this repository (gated behind trust_remote_code, default True in pyRadPlan.ml).