Bin Shizhen, Zhang Ji, Shen Luyao, Zhang Junjun, Wang Qi
Radiotherapy Center, Third Xiangya Hospital of Central South University, Changsha, China.
Radiotherapy Center, The Central Hospital of Shaoyang, Shaoyang, China.
Front Oncol. 2023 Jul 21;13:1094927. doi: 10.3389/fonc.2023.1094927. eCollection 2023.
To predict the gamma passing rate (GPR) in dosimetric verification of intensity-modulated radiotherapy (IMRT) using three machine learning models based on plan complexity and find the best prediction model by comparing and evaluating the prediction ability of the regression and classification models of three classical algorithms: artificial neural network (ANN), support vector machine (SVM) and random forest (RF).
269 clinical IMRT plans were chosen retrospectively and the GPRs of a total of 2340 fields by the 2%/2mm standard at the threshold of 10% were collected for dosimetric verification using electronic portal imaging device (EPID). Subsequently, the plan complexity feature values of each field were extracted and calculated, and a total of 6 machine learning models (classification and regression models for three algorithms) were trained to learn the relation between 21 plan complexity features and GPRs. Each model was optimized by tuning the hyperparameters and ten-fold cross validation. Finally, the GPRs predicted by the model were compared with measured values to verify the accuracy of the model, and the evaluation indicators were applied to evaluate each model to find the algorithm with the best prediction performance.
The RF algorithm had the optimal prediction effect on GPR, and its mean absolute error (MAE) on the test set was 1.81%, root mean squared error (RMSE) was 2.14%, and correlation coefficient (CC) was 0.72; SVM was the second and ANN was the worst. Among the classification models, the RF algorithm also had the optimal prediction performance with the highest area under the curve (AUC) value of 0.80, specificity and sensitivity of 0.80 and 0.68 respectively, followed by SVM and the worst ANN.
All the three classic algorithms, ANN, SVM, and RF, could realize the prediction and classification of GPR. The RF model based on plan complexity had the optimal prediction performance which could save valuable time for quality control workers to improve quality control efficiency.
使用基于计划复杂度的三种机器学习模型预测调强放射治疗（IMRT）剂量验证中的伽马通过率（GPR），并通过比较和评估三种经典算法（人工神经网络（ANN）、支持向量机（SVM）和随机森林（RF））的回归和分类模型的预测能力，找出最佳预测模型。
回顾性选取269个临床IMRT计划，使用电子射野影像装置（EPID）收集总共2340个射野在10%阈值下按2%/2mm标准的GPR，用于剂量验证。随后，提取并计算每个射野的计划复杂度特征值，训练总共6个机器学习模型（三种算法的分类和回归模型），以学习21个计划复杂度特征与GPR之间的关系。通过调整超参数和十折交叉验证对每个模型进行优化。最后，将模型预测的GPR与测量值进行比较，验证模型的准确性，并应用评估指标评估每个模型，以找出预测性能最佳的算法。
RF算法对GPR的预测效果最佳，其在测试集上的平均绝对误差（MAE）为1.81%，均方根误差（RMSE）为2.14%，相关系数（CC）为0.72；SVM次之；ANN最差。在分类模型中，RF算法的预测性能也最佳，曲线下面积（AUC）值最高，为0.80，特异性和敏感性分别为0.80和0.68，其次是SVM，ANN最差。
ANN、SVM和RF这三种经典算法均能实现GPR的预测和分类。基于计划复杂度的RF模型具有最佳的预测性能，可为质量控制人员节省宝贵时间，提高质量控制效率。
