Author: Yabo Fu, Yang Lei, Yu Lei, Haibo Lin, Ruirui Liu, Tian Liu, Kenneth Rosenzweig, Charles B. Simone, Shouyi Wei, Jiahan Zhang 👨🔬
Affiliation: Icahn School of Medicine at Mount Sinai, University of Nebraska Medical Center, Department of Medical Physics, Memorial Sloan Kettering Cancer Center, New York Proton Center 🌍
Purpose: Beam orientation optimization (BOO) in intensity-modulated radiation therapy (IMRT) is traditionally a complex, non-convex problem tackled with heuristic methods. This study benchmarks global optimal solutions for BOO, offering a reference for evaluating and improving heuristic approaches, thus bridging the gap between theoretical optimality and clinical feasibility.
Methods: This study developed a novel approach incorporating SOCP relaxation, sparse mixed integer programming (SMIP), and deep inverse optimization. The nonconvex dose-volume constraints were addressed using SOCP relaxation, replacing the ill-posed constraint with an SOCP constraint to maintain sparsity while ensuring convexity. BOO was framed as a SMIP problem, where binary variables indicated beam selection, and the problem was solved using an augmented Lagrange method. To accelerate optimization, a neural network, learning to approximate optimal solutions, optimized the iterative process and improved computational efficiency to 8× times. The study retrospectively analyzed 7 consecutive patients with locally advanced non-small cell lung cancer (NSCLC) (prescription dose of 60Gy), comparing automated BOO-selected beam angles with expert-selected ones. Dosimetric metrics, including PTV maximum dose and D98%, lung V20 and mean dose, and heart and esophagus mean dose were assessed.
Results: The automated BOO demonstrated superior dose conformity and sparing of critical structures. Specifically, the BOO plans achieved comparable PTV coverage (maximum: 61.7±1.4Gy vs. 61.9±1.3Gy, D98%: 59.5±0.6Gy vs. 59.6±0.5Gy) but demonstrated improved sparing for lungs (mean: 6.7±1.8Gy vs. 8.3±2.1Gy, V20: 11.8±2.0% vs. 15.3±2.5%), heart (mean: 5.3±1.2Gy vs. 6.8±1.4Gy), and esophagus (mean: 10.2±2.0Gy vs. 12.7±2.3Gy) compared to human-selected plans. This approach highlighted the potential of BOO to enhance treatment efficacy by optimizing beam angles more effectively than manual selection.
Conclusion: This framework establishes a benchmark for BOO in IMRT, enhancing heuristic methods through hybrid optimal solutions. The integration of SMIP and deep inverse optimization significantly improves computational efficiency and plan quality.