KEY POINTS
- This retrospective technical study developed a 3D Attention U-Net to generate intensity-modulated proton therapy spot weights directly from anatomical and dosimetric inputs, avoiding a conventional iterative dose-mimicking step. Training data comprised 95 prostate, 80 brain tumor and 100 lung cancer patients, yielding 380 beam samples per disease site, divided into 300 training, 40 validation and 40 independent-test samples with patient-level separation.
- The model received four input channels: optimized dose distribution, reference dose, CT image and spot-position mask. Training used robustly optimized single-field optimization plans with ±3.5% range and ±3 mm setup uncertainty, a 2-mm dose grid, and a nominal prescription of 2 Gy per fraction; robustness constraints were applied to targets but not to organs at risk.
- In independent testing, mean whole-body 2%/2 mm gamma passing rates were 99.73 ± 0.32% for prostate, 98.95 ± 0.86% for brain, and 92.75 ± 3.52% for lung. Corresponding target-volume passing rates were 98.16%, 94.83% and 74.03%, demonstrating substantially poorer performance in heterogeneous thoracic anatomy despite favorable global dose agreement in the other sites.
- At clinically scaled prescription levels, target D95 deviations were approximately 0.31 Gy for prostate and 0.75 Gy for brain, compared with 4.04 Gy for lung. The lung model demonstrated systematic target underdosage, with target mean-dose deviation of 1.29 Gy and differences in robust-scenario D98 reaching 3.33–4.79 Gy, making its current accuracy unsuitable for clinical lung planning.
- Spot-weight prediction was computationally fast. On a standard workstation with an RTX 3060 GPU, preparing the model inputs required approximately 15 seconds, followed by 4 seconds for spot-weight generation, giving a total runtime of less than 20 seconds for this module. This does not include the complete clinical workflow for imaging, segmentation, dose verification and plan approval.
- To investigate deliverability, the authors transferred model-generated weights into Eclipse and physically delivered one additional plan per disease site on a Varian ProBeam 360° system. Patient-specific QA passing rates at 3%/2 mm were 99.4% for prostate, 100% for brain and 99.3% for lung, all exceeding the institutional 95% criterion; successful QA of these three selected plans does not negate the broader lung target-dose errors.
- The framework remains limited to single-field optimization, with no multi-field optimization support, no explicit OAR robustness evaluation, and no integrated automatic safety check for clinically unacceptable predictions. The model was not prospectively deployed for online adaptation, and its performance depended strongly on anatomical site and tissue heterogeneity.
CLINICAL TAKEAWAY
Direct prediction of proton spot weights could substantially shorten the optimization component of adaptive IMPT workflows, particularly in anatomically homogeneous treatment sites. However, lung target-dose errors remain clinically unacceptable, and the method requires downstream dose verification, OAR robustness assessment and workflow integration before it can replace conventional optimization.