Comparing local and global optimization reveals untapped hybrid potentials for machine learning-based inverse planning

Abstract

Machine learning (ML) models for normal tissue complication probability (NTCP) can be embedded directly into inverse treatment planning, but the resulting optimization problems are generally nonconvex, and the choice between global and local optimization involves a trade-off between optimality guarantees and computational cost. Two open-source frameworks for ML-based inverse planning were compared: OptiCL (global mixed-integer optimization via constraint learning) and pyanno4rt (local nonlinear optimization via end-to-end differentiation). Logistic regression, linear support vector machine, and feed-forward neural network NTCP models for grade 2+ xerostomia and grade 2+ radiation pneumonitis were embedded. Plan optimization was performed for eight head-and-neck (HN) photon cases and seven lung proton cases. The plans were compared by target coverage, OAR dose metrics, spatial dose agreement, post-hoc predicted NTCP, and runtime. A projection-based gradient approximation was introduced to embed decision trees. In addition to that, warm-starting, constraint strengthening, and pruning were investigated as hybrid strategies for combining local and global optimization. Comparing OptiCL and pyanno4rt across the HN and lung treatment plans revealed broadly consistent dose distributions. Mean differences (OptiCL}$-$pyanno4rt [95% CI]) were small: for HN cases, -0.16 Gy [-0.37, 0.06] (parotid) and -0.00 Gy [-0.01, 0.00] (PTV); for lung cases, -0.52 Gy [-1.10, 0.06] (total lung) and -0.04 Gy [-0.07, -0.01] (PTV). Mean NTCP differences amounted to -0.08 p.p. [-0.49, 0.33] (HN) and -1.73 p.p. [-2.68, -0.77] (lung). OptiCL tended toward lower OAR doses and predicted NTCP, but required higher wall-clock runtime. Hybridization of the decision tree-based HN plan eliminated 80% of binary decision variables without compromising plan quality.