3032 - Deep Beam SBRT: An AI-Driven, Beam-Guided Automation Platform for Reproducible and High-Precision Lung SBRT Planning
Presenter(s)
R. Kaur1, E. Tajik-Mansoury2, J. N. Lee1, S. Patel3, N. Ju3, B. Floreza4, and R. Farjam1; 1Dept. of Radiation Oncology and Molecular Radiation Sciences, Johns Hopkins University School of Medicine, Baltimore, MD, 2Department of Radiation Oncology and Molecular Radiation Sciences, School of Medicine, Johns Hopkins University, Baltimore, MD, 3John Hopkins University, Washington, DC, 4Department of Radiation Oncology and Molecular Radiation Sciences, Johns Hopkins Medicine, Washington, DC
Purpose/Objective(s): To develop a beam-guided, AI-driven lung SBRT planning framework that integrates beam configuration, dose prediction, and optimization using curated beam-optimized clinical plans within an adaptive, continuously improving learning system.
Materials/Methods: Beam optimization was performed by maximizing therapeutic gain through minimizing the average lesion depth in the beam’s-eye-view (BEV). A curated repository of clinically approved lung SBRT plans generated in Raystation and by experienced dosimetrists was assembled for model development. A U-Net model was trained to predict the final clinical dose distribution, using the approved dose as ground truth. Model inputs included patient CT images, organ-at-risk (OAR) structure maps, and a 3D conformal beam-based dose distribution representing the optimized beam geometry to guide spatial dose learning. Ring structures and a prioritized structure map encoding distance-to-PTV and proximity-to-OARs were combined to generate a voxel-wise weighting map. Higher weights were assigned to the PTV and critical OARs (e.g., lungs, esophagus, pericardium). Training used a weighted mean squared error loss. The platform operates as an adaptive learning system. For each new patient, the predicted dose serves as an initialization for inverse planning objective definition. After physician and physics approval, the finalized plan is incorporated into the training dataset for iterative model updating. Initial training is intended to include up to 100 plans. Model performance was initially evaluated using leave-one-out cross-validation.
Results: The proposed BEV-based beam optimization reduced the average lesion depth by more than %20 (ranged: 3.66% to 45.78%). This geometric improvement translated into an average of 6.25% (p < 10-9), 15.03% (p < 10-9), and 9.86% (p < 10-7) decrease in mean lung dose, lung V10 and V5 respectively with similar target coverage and conformity index (p < 0.13). The subsequent beam-guided U-Net achieved stable convergence within 125 epochs. Across leave-one-out cross-validation, the mean voxel-wise MAE was 1.17 ± 0.65 Gy. Normalizing PTV to D95%, PTV mean dose differed from clinical plans by < ~5%, while lung V20 showed a deviation of nearly ~1.5%. The average 2%/2mm gamma passing rate was > ~97.5%.
Conclusion: We present a beam-guided automated planning framework for lung SBRT that integrates BEV-based beam optimization with deep learning–based dose prediction. The model accurately reproduced clinically approved dose distributions and demonstrated improved high-dose accuracy when beam information was incorporated. The adaptive learning strategy enables continuous refinement as new plans are added. This approach supports the feasibility of beam-informed, standardized SBRT planning. Additional validation warrants the utility of this system for lung SBRT planning workflow.