Physics-informed machine learning-based dose calculation for patients undergoing fluoroscopically guided...

To develop and evaluate a physics-informed machine-learning framework for rapid organ-dose estimation in fluoroscopically guided interventional procedures without requiring patient-specific imaging. Approach: We used 385 adult radiation dose structured reports from three procedure types-bilateral carotid angiography, percutaneous biliary drainage, and liver chemoembolization-comprising more than 50,000 irradiation events. A body-size-matched computational phantom was selected using patient sex, height, and weight. RDSR-derived beam parameters and engineered descriptors of beam-organ geometry and attenuation were used to train organ-specific XGBoost models for localized organs and a multi-output multilayer perceptron model for four distributed tissues. Monte Carlo-derived dose coefficients from NCIRF3 were used as target values. Main results: In the held-out test set of 12,683 irradiation events, 49.4% and 84.8% of XGBoost-based localized-organ estimates were within ±5% and ±20% of the Monte Carlo-derived values, respectively. The corresponding proportions were 88.2% and 98.0% for the distributed-tissue estimates and 87.6% and 98.3% for detriment-weighted dose. The median relative prediction error across organs and tissues was 0.37% (interquartile range, -3.68% to 4.99%). A reliability score was developed to support uncertainty-aware interpretation. The framework required approximately 0.3 s per irradiation event, corresponding to an approximately 290-fold acceleration relative to conventional Monte Carlo simulation under the benchmark conditions. Significance: The proposed image-free framework enables rapid organ-dose estimation when patient imaging is unavailable and may support large retrospective epidemiological studies and uncertainty-aware dose monitoring. Because development and validation were limited to three procedure types and relied on reconstructed phantom-beam geometry, external validation is required before application to other procedures or high-precision individual clinical decisions.