基于记忆学习法的放疗中呼吸运动预测技术的研究  

The Study on Predicting Respiratory Motion via Memory-Based Learning in Radiotherapy

作　　者：万伟权 张慧连 徐子海[3] 贺志强[1] 陈超敏[1] 

机构地区：[1]南方医科大学生物医学工程学院,广州510515 [2]广东新华软件外包有限公司,广州510515 [3]解放军303医院放射治疗中心,南宁530021

出　　处：《中国生物医学工程学报》2014年第2期148-154,共7页Chinese Journal of Biomedical Engineering

基　　金：广东省重大科技专项(2012A080104010)

摘　　要：对胸腹部肿瘤进行实时跟踪放疗时,需要通过预测来补偿系统延迟.然而,由于呼吸运动的复杂性,传统方法难以满足要求.本文应用一种基于记忆学习法进行呼吸预测,该方法首先存储训练数据到记忆中,然后查找相关数据应答当前查询.在此基础上,采用“滑窗法”动态更新训练数据集,并针对预测过程中出现的“病态矩阵”采用脊回归进一步改进算法,使算法的精确性和鲁棒性有了很大提高.实验使用POLARIS红外定位系统采集了10例正常人体表的红外反射标记物的呼吸运动数据样本,平均幅度约为20 mm（9.2 ～37.8 mm）,采用改进后的基于记忆学习法（预测步长为1 s）,平均绝对误差约为0.3 mm（0.08～0.8 mm）,每次估值耗时约1 ms.所提出的方法能够准确和实时捕捉复杂的呼吸运动轨迹.Prediction is necessary to compensate the system latency in the real-time tracking radiation therapy for thoracic and abdominal cancers. However, because of the complexity of the breathing motion, conventional methods are far from clinical requirements. This paper proposed a memory-based learning method to predict respiratory motion. The method stores the training data in memory, then finds relevant data to answer a particular query. Furthermore, the paper adopts dynamic update the training data method and ridge regression aimed at ＂ill-condition matrix＂ to greatly improve the accuracy and robustness of the algorithm. Our experiment collected ten respiratory motion data with average amplitude of 20 mm （9.2 - 37.8 mm） from humans＇ body surface using POLARIS infrared positioning system. Using our methods （prediction horizon is Is）, mean absolute error （MAE） was reduced to 0.3 mm （0. 08 - 0.8 mm） , per estimate takes 1 ms. The results confirm that the proposed method is able to capture highly complex breathing movement accurately in real time.

关 键 词：放射治疗 记忆学习法 呼吸运动 实时 预测 

分 类 号：R814[医药卫生—影像医学与核医学]