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.
Related
- From Energy-Force Weighting to Primal-Dual Optimization of Machine-Learned Interatomic Potentials
- AI-Accelerated Gyrokinetic Predictions of Turbulent Transport for Stellarator Design Optimization and Experimental Planning
- In-context Learning of Single-index Targets: Comparing Kernel and Feature Learners
- Optimal Bias Potentials via Ergodic Optimal Control and Generator Learning
- Benchmarking machine-learning interatomic potentials using classical force fields
BibTeX
@misc{ortkamp2026comparing,
title = {{Comparing local and global optimization reveals untapped hybrid potentials for machine learning-based inverse planning}},
author = {Tim Ortkamp and Florian Leininger and Zhongxing Liao and Niklas Wahl and Ali Ajdari},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.30424},
eprint = {2610.30424},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.30424},
}
RIS
TY - UNPB
TI - Comparing local and global optimization reveals untapped hybrid potentials for machine learning-based inverse planning
AU - Tim Ortkamp
AU - Florian Leininger
AU - Zhongxing Liao
AU - Niklas Wahl
AU - Ali Ajdari
PY - 2026
DA - 2026/10/21
PB - arXiv
DO - 10.48550/arxiv.2610.30424
UR - https://arxiv.org/abs/2610.30424
AB - 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.
ER -