Optimization of Particle Source Dose Planning Algorithm Based on Neural Network

• 1、School of Radiation Medicine and Protection,Soochow University,Suzhou,215123

Abstract：Brachytherapy with radioactive seeds implantation is a crucial approach in precision radiotherapy for malignant tumors. In recent years, neural network technology has demonstrated remarkable potential in medical image analysis and dose prediction. However, existing research primarily focuses on predicting dose distribution, lacking a comprehensive planning process for seed placement and dose optimization. To address this issue, this paper centers on the automatic placement of seed sources and dose optimization, constructing an intelligent dose planning algorithm based on neural networks and optimizing clinical application processes. Firstly, analyze the dose calculation mechanism and clinical evaluation indicators of particle implantation therapy, and establish a physical and evaluation model for particle source dose planning. Secondly, construct a dose prediction model based on an improved Attention U-Net, integrating CT images, target contour, and particle parameters to achieve high-precision prediction of three-dimensional dose distribution. Subsequently, propose an inverse source placement algorithm combining simulated annealing and DVH evaluation, which combines predicted dose distribution with clinical dose constraints to optimize particle position and quantity. Finally, construct a complete intelligent dose planning process, and validate the algorithm performance and application effectiveness through clinical case data. The experimental results showed that the target volume V100 increased to 92.8%, the dose homogeneity index HI decreased to 1.05, and the overdose rate of critical organs decreased from 11.1% to 2.2%. The reasoning time for a single patient was less than 1.18 minutes, which was more than 95% shorter than traditional manual planning. In the verification of prostate cancer cases, the plan automatically generated by the algorithm met the requirements of V100≥92.8% and D90≥121.2Gy, with all doses to critical organs within the safety threshold. This study proposes a neural network-based particle source dose planning method tailored for clinical applications, establishing a comprehensive automatic planning workflow. This approach effectively enhances the accuracy of dose prediction and planning efficiency, ensuring target dose coverage while reducing the radiation dose to critical organs. The planning time is significantly shortened, and the generated dose distribution meets the requirements of clinical Dose Volume Histogram (DVH) evaluation indicators. This provides a new technological path for the intelligent development of brachytherapy, and holds significant importance in improving the quality of treatment plans, reducing planning costs, and promoting the standardization of precision radiotherapy.

Keywords： neural network particle source dose planning brachytherapy simulated annealing