Author: Somayeh Gholami, Saeedeh Ketabi, Jeremy Kunz, Ali Yousefi 👨🔬
Affiliation: Department of Radiation Oncology, University of Utah, University of Utah, Department of Management- Operations Research, University of Isfahan 🌍
Purpose: This paper aims to present a novel approach for automatic knowledge-based planning optimization for Intensity Modulated Radiotherapy, along with the two downsizing techniques for improving the computational efficiency and reducing the solving time.
Methods: Two mathematical models have been applied to optimizing the treatment plan, and auto-imputing weights of objective function. QuadLin model for treatment plan optimization and the other model for automatically adjusting the weights of the QuadLin objective function. The study emphasizes improving computational efficiency and reducing solving time by introducing an innovative algorithm, SVSIDB, which clusters voxels based on the dominant beamlet concept. Additionally, the hybrid ultra-heuristic ABC-K-Means technique was developed for voxel clustering. Finally, more efficient pipeline has been selected according to the results. A generally published real dataset of 100 head and neck patients, Open-KBP, has been used in the current work.
Results: The weights of the objective function have been adjusted automatically by the latter mathematical model. The performance of two proposed clustering techniques was remarkable regarding to reducing the solving time. SVSIDB as a systematic voxel clustering method reduced solving time by nearly 50% compared to Full-data models, while maintaining treatment plan quality. SVSIDB achieved an 81.3% clinical criteria satisfaction rate, which is 10% higher than ABC-K-Means
Conclusion: in this study a novel KBP-based planning method was developed which is optimized, automatic, and efficient. An innovative heuristic downsizing algorithm called SVSIDB was proposed which reduced solving time by 50%. Notably, the SVSIDB-QuadLin pipeline not only reduced solving time but also enhanced plan quality by an average of 12% in satisfying clinical criteria, surpassing models based on Full data, marking a significant advancement over prior research.