基于神经网络的非晶硅EPID剂量标定方法研究

发布时间：2018-05-01 21:36

本文选题：调强放射治疗 + 神经网络　；参考：《河南工业大学》2017年硕士论文

【摘要】：随着物理学与生物科学理论的不断完善以及放射治疗技术的快速发展,放射治疗已经成为目前治疗肿瘤的主要手段之一。为了提高治疗效果,降低射线对人体的副作用,在对病人实施放射治疗时,需要精准地控制照射野的剂量分布,同时尽可能降低甚至避免放射线对靶区附近健康组织和器官的伤害。由于调强放射治疗方式的射野复杂度和剂量率比其他放射治疗方式的射野复杂度和剂量率高,治疗实施的各个环节均有可能产生误差,因此需要对调强放射治疗计划进行剂量验证。基于a-Si EPID的剂量验证是目前具有良好发展前途的剂量验证技术之一。本论文的工作就是应用计算机技术,研究a-Si EPID子野的剂量响应特性,探讨a-Si EPID灰度影像剂量标定的方法,为剂量验证等后续工作提供技术支撑。本论文首先采集了数字医用直线加速器在不同输出剂量情况下a-Si EPID装置的灰度影像,并同时采用三维水箱采集了照射野输出剂量的空间分布数据。由于a-Si EPID灰度影像的空间分辨率远高于采集到的剂量数据的分辨率,为了获得充足的数据,需要对采集的文本格式的低分辨率剂量分布数据进行数据结构转换和数据补充。数据采集过程中,往往有很多因素会影响数据采集的精确度,而神经网络具有较强的自学习能力、处理不精确信息的能力、较强的抗干扰和抗噪声能力,本文采用神经网络来处理放射治疗学的数据。在研究分析各种插值、拟合算法特点的基础上,根据照射野剂量分布的特点,采用三次样条插值法对剂量平稳区域进行插值,采用基于遗传算法的广义回归神经网络模型对剂量分布复杂的区域进行拟合,最终形成2048×2048的二维矩阵,并采用BP神经网络模型研究非晶硅EPID子野的剂量响应特性。对1-12MU影像按射线剂量的大小顺序进行融合,鉴于非晶硅EPID子野剂量响应特性的分散性和1-12MU融合影像的所具有的空间特征,设计并建立空间线性神经网络模型,使用验证样本评价该模型,表明采用该模型来实现剂量标定的可行性。
[Abstract]:With the continuous improvement of physics and biological science theory and the rapid development of radiotherapy technology, radiotherapy has become one of the main methods of tumor treatment. In order to improve the therapeutic effect and reduce the side effects of radiation on the human body, it is necessary to accurately control the dose distribution of the radiation field during the radiotherapy of patients. At the same time minimize or even avoid radiation damage to healthy tissues and organs near the target area. Because the radiation field complexity and dose rate of intensity modulated radiation therapy are higher than those of other radiation therapy methods, errors may occur in all aspects of the treatment. It is therefore necessary to verify the dose of the IMRT program. Dose verification based on a-Si EPID is one of the promising dose verification techniques. The work of this paper is to use computer technology to study the dose response characteristics of a-Si EPID subfield, to discuss the method of dose calibration for a-Si EPID gray image, and to provide technical support for further work such as dose verification. In this paper, the grayscale images of the a-Si EPID device under different output doses of the digital medical linear accelerator are first collected, and the spatial distribution data of the output dose of the irradiation field are also collected by using the three-dimensional water tank. Because the spatial resolution of a-Si EPID gray image is much higher than that of the collected dose data, in order to obtain sufficient data, it is necessary to transform and supplement the collected low-resolution dose distribution data in text format. In the process of data acquisition, there are many factors which will affect the accuracy of data acquisition, but the neural network has strong ability of self-learning, dealing with imprecise information, strong anti-interference and anti-noise ability. Neural networks are used to process radiotherapy data. On the basis of studying and analyzing the characteristics of various interpolation and fitting algorithms, according to the characteristics of radiation field dose distribution, the cubic spline interpolation method is used to interpolate the steady region of dose. The generalized regression neural network model based on genetic algorithm was used to fit the complex region of dose distribution, and a two-dimensional matrix of 2048 脳 2048 was formed. The dose response characteristics of EPID subfield in amorphous silicon were studied by BP neural network model. The 1-12MU images are fused in the order of radiation dose. In view of the dispersion of the field dose response characteristics of amorphous silicon EPID and the spatial characteristics of 1-12MU fusion images, a spatial linear neural network model is designed and established. Validation samples are used to evaluate the model, which shows the feasibility of using the model to achieve dose calibration.
【学位授予单位】：河南工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R730.55;TP183

【参考文献】