Deep learning-based spot weight generation for online adaptive IMPT: implementation and technical feasibility on a clinical delivery system.

PURPOSE: To investigate the feasibility of a deep learning-based method for generating intensity-modulated proton therapy (IMPT) spot weights to enable a faster, more automated workflow. METHODS: A 3D Attention U-Net with a masked mean absolute error loss generated spot weights from four input channels (optimized plan dose, reference dose, CT image, spot mask), using planned spot weights as labels, and trained and tested on prostate, brain, and lung cases. Robustness was evaluated on target dose metrics only. The workflow was validated in the Eclipse® treatment planning system and delivered on a ProBeam 360° system. RESULTS: For weights in the range 0-5, mean absolute errors were below 0.8 for the prostate and brain models (15.9% and 48.1% of all weights, respectively) and below 1.7 for the lung model (31.2%). Mean relative errors were below 10% for the prostate model (5-65, 83.7%), 20% for the brain model (5-65, 51.7%), and 40% for the lung model (5-110, 68.6%); the remaining weights lay above these ranges and accounted for <0.5% of all weights. Gamma passing rates within the body contours (2%, 2 mm, 10% dose threshold) were 99.73 ± 0.32%, 98.95 ± 0.86%, and 92.75 ± 3.52% for the prostate, brain, and lung models, respectively. All delivered plans met the clinical QA acceptance criteria. CONCLUSION: The proposed framework demonstrates the technical feasibility of generating IMPT spot weights with deep learning and implementing them on a clinical delivery system. Further improvement is needed for heterogeneous sites such as the lung.