实时跟踪放疗中关联模型和预测算法

Review of Correspondence Models and Prediction Algorithms in Real-time Tracking Radiotherapy

目的:对胸腹部肿瘤进行图像引导实时跟踪放射治疗时,通常难以直接监控肿瘤或其他内部解剖结构的呼吸运动轨迹,因此,利用体外信号获取内部肿瘤运动信息是很好的替代方法。方法:首先同步采集体内数据和体外数据,建立关联模型,然后采集体外数据拟合关联模型得到体内信号的估值;由于系统延迟的存在,还必须通过呼吸运动预测算法进行补偿。关联模型主要分为直接关联模型和间接关联模型,预测算法可以分为基于模型的预测算法以及无模型的预测算法。结果:直接关联模型把体外运动信号与内部肿瘤运动信号的相关性直接定义为一个函数。间接关联模型并不直接定义内-外相关性,而是利用一些内变量参数化呼吸运动模型,同时估计体外信号和体内信号的值,在获得真实的体外数据时,最优化内变量使得体外信号的估计值与真实值最匹配。基于模型预测算法对呼吸运动信号进行预测通常建立在平稳性和周期性的假设上,但该假设可能是错误的。无模型的预测算法更具优势的是不需要预先了解呼吸运动信号。结论:目前绝大部分的关联模型和预测算法在有限的实验或实际数据样本上都提升了某方面的性能,但都只限于文献报道,距离临床应用还需大量真实数据的验证。

Objective To real-time tracking the cancers in thoracic and abdominal, it is often difficult to directly monitor the res- piratory motion of the tumor and other internal anatomy, therefore, an alternative option is to obtain the internal motion using the surrogate singal （s）. Methods Firstly, imaging data was simultaneously acquired with the surrogates to build a correspon- dence model, and then the internal motion could be estimated from the surrogate data. Respiratory motion prediction was also necessary to compensate the system latency. There were two types of correspondence model： direct correspondence models and indirect correspondence models. And there were two categories of prediction algorithms： model-based prediction and model- free prediction. Results A direct correspondence model estimated the motion as a direct function of the surrogate signal （s）, while an indirect correspondence model did not directly relate the motion to the surrogate data. Instead, they parameterised the motion using one or more internal variables, and made estimates of the surrogate data as well as the motion data. When surro- gate data was acquired during a procedure the internal variables were optimized to give the best match between the estimated surrogate data and the measured surrogate data. Model-based prediction algorithms for the prediction of respiratory motion sig- nals were usually built on the assumption of stationarity and periodicity, but the assumption would be wrong. Model-free pre- diction methods had advantage of no need for knowledge about the respiratory motion signal. Conclusion So far, almost all the correspondence models and prediction algorithms can improve the performance of a certain aspect on the limited data samples, but just described in the literature, so need to be verified using clinically realistic data for clinical application.