Warm-IP: Fast Constrained Optimization for Automated and Adaptive Radiotherapy Treatment Planning
Abstract
Purpose
Fast and reliable optimization remains a major bottleneck in automated and adaptive radiotherapy treatment planning. Most commercial and research planning systems rely on gradient-based optimization methods, which can converge slowly near the optimal solution and provide limited support for enforcing strict (“hard”) clinical constraints beyond simple non-negativity. This limitation is particularly problematic for online-adaptive-planning, where essential clinical-criteria (e.g., maximum dose, DVH-constraints) must be guaranteed without manual parameter tuning. Although primal–dual interior-point methods are a gold standard for constrained optimization and naturally support hard constraints, their high computational cost has limited clinical adoption. We present Warm-IP, a fast interior-point optimization algorithm specifically designed for radiotherapy treatment planning.
Methods
Warm-IP achieves high performance through two key innovations. First, it employs an efficient warm-start using the Alternating Direction Method of Multipliers (ADMM), initializing the interior-point method close to the optimal solution and substantially reducing convergence time. Second, Warm-IP exploits problem structures unique to radiotherapy optimization. The most computationally intensive step of interior-point methods—solving large linear systems arising from Karush–Kuhn–Tucker (KKT) conditions—is accelerated by identifying large diagonal substructures that enable efficient back-substitution. Warm-IP was evaluated on IMRT optimization problems for 62 lung patients and compared against two commercial and three open-source solvers using the open-source PortPy platform and its publicly available datasets. All code will be released publicly to support reproducible, community-driven research.
Results
Warm-IP achieved approximately 2× faster solution times compared with the fastest commercial and open-source solvers on CPU. While interior-point methods are typically not GPU friendly, Warm-IP reformulates the linear system solves into smaller, GPU-efficient operations, yielding preliminary speedups of up to 5× on GPU.
Conclusion
Warm-IP enables fast, robust treatment planning optimization with native support for hard clinical constraints. Its performance and reliability make it well suited for automated and online adaptive radiotherapy planning under tight clinical time constraints.