Deep-Learning-Based Proton Spot Map Generation from Dose and Linear Energy Transfer
Abstract
Purpose
To develop and evaluate a deep-learning approach that rapidly reconstructs clinically deliverable proton spot maps (PSMs) directly from dose or linear energy transfer (LET) distributions, bypassing conventional iterative optimization and supporting adaptive and biologically driven proton therapy.
Methods
We defined a projected PSM (prPSM) representation by projecting each clinically delivered PSM onto the patient CT along the beam paths, encoding spot lateral coordinates, stopping power ratio, and water‑equivalent thickness. A SwinUNETR model was trained on 259 four‑field prostate SBRT pencil‑beam scanning proton plans to predict prPSM from either dose or LET. Among them, 80% for training, 10% for validation, and 10% for test. Separate models were trained for dose→prPSM and LET→prPSM. Predicted prPSMs were then decomposed via linear regression into deliverable PSMs, yielding spot positions, energies, and protons per spot. Plan fidelity was assessed by Monte Carlo dose calculations based on the reconstructed PSMs and comparison with the clinical plans using gamma analysis, mean absolute error (MAE), and dose-volume-histogram (DVH) metrics.
Results
The deep‑learning model predicted prPSMs in 0.02 s with a structural similarity index of 0.92±0.04 and voxel‑wise MAE of 0.05±0.03 (normalized units). Final PSMs were generated in approximately 4 s per field and closely reproduced the reference spot and energy distributions. Monte Carlo dose from the generated PSMs achieved a gamma passing rate of 92.4%±1.5% and a target‑region MAE of 0.54±0.1 Gy relative to the clinically delivered dose. For a representative case, target DVH metrics were D90=12.36 Gy (AI) versus 12.98 Gy (clinical) and D5=14.97 Gy (AI) versus 13.88 Gy (clinical). The LET‑based model demonstrated performance comparable to the dose‑based model.
Conclusion
This fast, data‑driven framework infers deliverable PSMs directly from dose or LET, removing the need for time‑consuming inverse planning and providing a practical building block for adaptive and biologically optimized proton therapy.