【作者】 张力；

【导师】 刘永；

【作者基本信息】 电子科技大学 ， 光学工程， 2020， 硕士

【摘要】 激光治疗是治疗糖尿病视网膜病变的主要手段之一。然而目前市面上的眼底激光治疗仪自动化程度不高,激光在眼底视网膜的定位和聚焦方面基本上都是依赖于医生手工操作,医生手工操作精度低,治疗效果差,严重时甚至会损害患者的视网膜,而且手工操作效率低,治疗时间长,会造成医生疲劳和患者不适的情况。针对上述问题,本文对眼底激光治疗智能辅助系统的关键技术进行深入研究。本文设计了眼底激光治疗智能辅助系统的总体设计方案,该系统主要由眼底数字图像采集单元、上位机软件控制单元、嵌入式激光控制单元组成,分析了各个模块在整个系统中的作用以及各模块之间的相互联系,详细分析了整个模块所选用到的硬件设备,详细介绍了各个硬件的详细参数信息。本文着重对眼底激光治疗智能辅助系统中的眼底特征识别算法进行研究,详细介绍了算法所涉及到的基础知识,并且基于双线性卷积神经网络结构,提出了双线性空洞卷积U-net神经网络模型算法,用于对眼底视网膜的中央凹区域和视盘区域的图像语义分割,详细分析了该神经网络模型的设计原理,阐述了该神经网络模型的搭建和实现过程,详细说明了眼底视网膜数据集的制作过程。本文基于振镜系统设计了嵌入式激光控制系统,详细说明了嵌入式激光控制系统的实现原理,推导了激光二维扫描的数学模型,并且以此设计了激光振镜扫描控制卡;设计了激光动态聚焦光学系统,详细计算了激光动态聚焦光学系统的光学传播模型,使用基于深度学习对激光进行动态聚焦。最后,我们对提出的双线性空洞卷积U-net神经网络进行测试,并将测试结果与常见的神经网络模型从准确率,运行速度和模型内存大小三个方面进行比较,发现本文提出的神经网络综合结果为最优。本文也对设计的嵌入式激光控制系统进行了实验环境搭建,控制软件设计,最后进行大量的实验测试,测试结果表明该嵌入式激光控制系统的定位精度,聚焦精度以及扫描速度可以大大满足实际医疗需求。更多还原

【Abstract】 Laser therapy is one of the main methods to treat diabetic retinopathy.However,at present,the automation of fundus laser therapy instrument is not high.The laser positioning and focusing of fundus retina are basically dependent on the manual operation of doctors.The accuracy of the manual operation of doctors is low,and the treatment effect is poor.In serious cases,it will even damage the patients’ retina.Moreover,the manual operation efficiency is low,and the treatment time is long,which will cause doctors’ fatigue and patients’ discomfort.In view of the above problems,the key technologies of the intelligent assistant system for fundus laser therapy are studied in this paper.First of all,this paper puts forward the overall design scheme of the intelligent assistant system of fundus laser therapy,which is mainly composed of fundus digital image acquisition unit,upper computer software control unit,embedded laser control unit.It analyzes the function of each module in the whole system and the relationship between each module,and analyzes in detail the hardware equipment selected in the whole module The detailed parameter information of each hardware is introduced in detail.Secondly,this paper focuses on the research of fundus feature recognition algorithm in the intelligent assistant system of fundus laser therapy,and introduces the basic knowledge involved in the algorithm in detail.Based on the structure of bilinear convolution neural network,a bilinear hole convolution U-net neural network model is proposed,which is used for image semantic segmentation of fovea and optic disc areas of fundus retina This paper analyzes the design principle of the neural network model,explains the construction and implementation process of the neural network model,and introduces the production process of the fundus retina data set in detail.Thirdly,based on the galvanometer system,the embedded laser control system is designed in this paper.The realization principle of the embedded laser control system is explained in detail.The mathematical model of two-dimensional laser scanning is deduced,and the laser galvanometer scanning control card is designed.The laser dynamic focusing optical system is designed,and the mathematical model of the laser dynamic focusing optical system is calculated in detail,using the method based on the figure The focusing depth method of image processing focuses the laser dynamically.Finally,we test the proposed neural network,and compare the test results with the common neural network model from three aspects: accuracy,running speed and model memory size.It is found that the comprehensive results of the neural network proposed in this paper are the best.In this paper,the design of the embedded laser control system is also carried out in the experimental environment,control software design,and finally a large number of experimental tests.The test results show that the positioning accuracy,focusing accuracy and scanning speed of the embedded laser control system can greatly meet the actual medical needs.更多还原