Preprint compares local and global optimization for machine-learning-based radiotherapy planning

Findings

Researchers compared two open-source frameworks that embed machine-learning models of normal-tissue complication probability into inverse treatment planning. OptiCL used global mixed-integer optimization through constraint learning, while pyanno4rt used local nonlinear optimization through end-to-end differentiation. The study included eight head-and-neck photon cases and seven lung proton cases.

The embedded models included logistic regression, linear support vector machines and feed-forward neural networks. They predicted grade 2 or worse xerostomia and grade 2 or worse radiation pneumonitis. Researchers assessed target coverage, organ-at-risk dose measures, spatial dose agreement, post-hoc predicted complication probabilities and wall-clock runtime.

The researchers reported broadly consistent dose distributions. For head-and-neck cases, mean OptiCL-minus-pyanno4rt differences were −0.16 Gy for the parotid, with a 95% confidence interval of −0.37 to 0.06, and approximately 0 Gy for the planning target volume. For lung cases, the corresponding differences were −0.52 Gy for total lung, with an interval of −1.10 to 0.06, and −0.04 Gy for the target volume, with an interval of −0.07 to −0.01.

Mean differences in predicted complication probability were −0.08 percentage points for head-and-neck cases, with a 95% confidence interval of −0.49 to 0.33, and −1.73 percentage points for lung cases, with an interval of −2.68 to −0.77. OptiCL tended toward lower organ-at-risk doses and predicted probabilities but required more wall-clock time.

Why it matters

The comparison highlights a potential trade-off: the tested approaches produced broadly consistent dose distributions, while the global framework took longer and tended toward somewhat lower organ-at-risk doses and predicted complication probabilities. Researchers also tested warm-starting, constraint strengthening and pruning as hybrid strategies. In one decision-tree-based head-and-neck plan, hybridization eliminated 80% of binary decision variables without compromising reported plan quality.

What remains uncertain

This study is a preprint and has not been peer reviewed. It evaluated 15 planning cases, and the abstract does not provide case-selection details, absolute runtimes, computing hardware or enough methodological information to assess robustness or generalizability. It also does not establish which hybrid component accounted for the reduction in binary variables or whether that result extends beyond the specific decision-tree plan.

The complication probabilities were machine-learning model outputs assessed after planning, not observed patient outcomes. The comparison therefore does not show that either framework reduces complications or improves health outcomes. Clinical utility and performance across broader sets of treatment plans cannot be assessed from the available abstract.