基于实时超声图像引导技术和线性判别模型分析前列腺癌放疗分次内运动模式

目的基于经会阴超声(TPUS)实时扫描技术和线性判别分析(LDA)法,定性分析并自动判别前列腺癌放疗分次内运动模式,为个体化精确放疗奠定基础。方法应用TPUS技术记录了61例前列腺癌患者共1265个分次近百万个实时监测数据,划分为稳定型、波浪型、小偏执型、银叉型、回归型、大偏执型和稽留型运动模式。对运动轨迹量化并提取特征参数,通过LDA法建立判别方程式,评估训练集和测试集的判别效果。结果平均每位患者存在4种不同的运动模式,不稳定型占(35.00±21.49)%。随着治疗次数的增加,运动轨迹并未表现出越来越稳定的趋势,不同模式的出现极不规则。构建的线性判别模型对训练集和测试集的判别准确率分别为90.4%和89.5%,敏感性和特异性分别为84.9%和91.1%。结论前列腺癌患者分次内运动模式多样且随机,具有不可预测的特点。LDA法可以有效地对分次内运动模式进行判别,同时在治疗过程中利用判别方程和中心坐标实现对未知样本的自动鉴别。

Auto-analysis intrafraction prostate movement patterns based on transperineal ultrasound real-time tracking system and linear discriminant model

Objective To qualitatively analyze and auto-discriminate intrafraction prostate movement patterns based on transperineal ultrasound (TPUS) technique and linear discriminant analysis (LDA) method, which lays a solid foundation for individualized precise radiotherapy. Methods A total of 1265 intrafraction motion trajectories and millions of monitoring data were recorded by TPUS from 61 prostate cancer patients. Seven typical patterns were considered:stable at baseline, slight transient excursion, persistent excursion and continuous drift, as well as obvious transient excursion, persistent excursion and continuous drift. Motion trajectory diagrams associated with the displacement-time function were generated by MATLAB programming. Valuable features were selected and different patterns were identified. Discriminant accuracy and receiver operating characteristic (ROC) curve were utilized to evaluate the performance of LDA model. Results Four movement patterns were found per patient during the whole treatment process, and unstable type was occupied at (35.00±21.49)%. With the increase of treatment times, motion trajectories did not show an increasingly stable trend, and the appearance of different patterns was extremely irregular. After quantitative analysis, discriminant accuracy of LDA method was 90.4% for the training set and 89.5% for the testing set, with a sensitivity of 84.9% and a specificity of 91.1%. Conclusion Intrafraction movement patterns are characterized by diversity and randomness. The LDA method can be used to discriminate different movement patterns effectively, and the unknown samples can be identified by discriminant equations and centroid coordinates.