A GPU-Based Multi-Criteria Optimization Algorithm for LDR Brachytherapy
Abstract
Purpose
In current prostate permanent implant (PPI) low-dose-rate (LDR) brachytherapy planning, algorithm-generated treatment plans are frequently sub-optimal, requiring time-consuming and user-dependent manual adjustments. This work aims to develop and evaluate a GPU-accelerated inverse treatment planning algorithm for LDR brachytherapy that reduces planning time while enabling systematic exploration of dosimetric trade-offs.
Methods
A GPU-accelerated simulated annealing algorithm was implemented within a multi-criteria optimization framework (gMCO-LDR) and benchmarked against a CPU-based implementation. Algorithm convergence behavior and Pareto front characteristics were analyzed. Plan quality was assessed using dose–volume histogram (DVH) indices and target coverage metrics. Performance was evaluated on ten post-implant clinical cases and one pre-implant ultrasound-based case including a dominant intraprostatic lesion (DIL). The pre-implant case was compared to a physician-approved clinical plan using institutional DVH objectives. Mean planning time was measured for the generation of 1000 optimized plans.
Results
For all ten post-implant clinical cases, gMCO-LDR generated multiple clinically acceptable treatment plans. Using a stopping criterion of 200,000 iterations, the GPU-accelerated algorithm generated 1000 optimized plans in a mean time of 3 minutes across all 11 cases. The GPU implementation achieved an average computational speedup factor of 24 compared to the CPU-based algorithm when generating large plan populations. In the pre-implant DIL case, the optimized plan achieved dosimetric quality comparable to the clinical plan, with a 1% increase in prostate target coverage (V100), while maintaining clinically acceptable DIL boost coverage (V150>95%) and urethral sparing (V150=0%).
Conclusion
The proposed GPU-accelerated simulated annealing algorithm enables the generation of hundreds of Pareto-optimal LDR brachytherapy plans within clinically viable timeframes. By reducing user dependence and facilitating systematic exploration of plan trade-offs, gMCO-LDR addresses key limitations of current PPI planning workflows. Coupled with an adapted graphical user interface, this approach represents a practical solution for inverse LDR brachytherapy treatment planning.