【作者】 李东升；

【导师】 陈春晓；

【作者基本信息】 南京航空航天大学 ， 生物医学工程， 2020， 硕士

【摘要】 三维荧光分子成像可以对成像组织内的荧光区域进行定位、定量和持续监测,是一种极具潜力的分子影像技术。随着三维荧光分子成像技术的不断发展,其在肿瘤研究、基因表达和药物开发等研究中广受关注。三维荧光分子成像技术缺乏必要的解剖结构信息,但CT成像可以提供较为丰富的组织结构信息。结合CT成像的多模态三维荧光分子成像逐渐成为三维荧光分子成像技术的重点研究内容。然而,受制于CT数据的采集方式,CT体数据中往往含有扫描床板的干扰。合理恰当的体数据预处理方式可以为后续操作提供良好的原始数据,对提高光源重建精度有重要意义。此外,逆向光源重建问题是三维荧光分子成像技术的核心内容。由于可测量数据有限及生物组织内部的光子传输过程比较复杂,使得逆向光源重建存在严重的病态性。传统三维荧光分子成像多采用正则化方式进行求解,其重建精度和速度还有待进一步提高。本文从深度学习的角度出发,对三维荧光分子成像方法进行了改进。本文主要工作如下:(1)研究了三维荧光分子成像系统的理论基础,重点研究了前向光子传输模型、逆向光源重建算法和CT图像预处理方法;(2)针对传统CT数据预处理方法自动分割程度低、普适性差的问题,引入深度学习语义分割网络模型进行体数据分割。从特征融合和减少特征语义差异的角度对U-Net深度学习模型进行改进,提出了密集门限网络(DG-Net)。CT图像预处理实验表明,本文提出的方法可以有效提高CT图像分割精度。(3)针对传统逆向光源重建算法重建速度较慢,重建精度不足的问题,本文将深度学习方法应用到逆向光源重建过程中,根据三维荧光分子成像的数据结构特点,提出使用图卷积构建光源重建预测网络。逆向光源重建实验表明,图卷积网络可以快速准确地进行光源重建,且具有较高的鲁棒性。(4)在研究成果的基础上对课题组搭建的小动物光学成像系统进行改进。将本文提出的深度学习模型部署到系统的CT图像预处理和光源重建模块中,并开展实验验证算法性能。实验表明,相对于传统算法,本文提出的算法更加准确。更多还原

【Abstract】 Three-dimensional fluorescence molecular tomography is a promising molecular imaging technique for positioning,quantizing and monitoring fluorescent region in tissue.With the continuous development of three-dimensional fluorescence molecular tomography,it has attracted much attention in tumor research,gene express and drug development.Fluorescence molecular tomography lacks the necessary anatomical information,while CT imaging can provide spatial information.The multi-model fluorescence molecular tomography technology gradually becomes the focus of fluorescence molecular tomography technology research.However,subject to CT data acquisition methods,CT volume data often contains scanning bed which has an adverse effect on the subsequent process.Reasonable and appropriate volume preprocessing could provide original data for subsequent algorithms,and is significant for promoting the accuracy of fluorescent source reconstruction.Meanwhile,fluorescent source reconstruction is the core of fluorescence molecular tomography.However,due to the limited measurement data and the complex photon transfer process in tissues,the problem of fluorescent source reconstruction is seriously ill-conditioned.Traditional three-dimensional fluorescence molecular tomography is mostly solved by regularization.The accuracy and speed of the regularization method need to be further promoted.In this paper,we improved the three dimensional fluorescence molecular tomography from the perspective of deep learning.The main work of this paper is as follows:(1)The theoretical basis of the three-dimensional fluorescence molecular imaging system is studied.Among them,the forward photon transmission model,the inverse light source reconstruction algorithm and the CT image preprocessing method is mainly studied.(2)To solve the problem of low automatic and poor universality in traditional CT data processing,the deep learning semantic segmentation network is introduced to the CT volume data segmentation process.And from the perspective of feature fusion and semantic information discrepancy reduction,a dense gate network(DG-Net)is proposed to improve the performance of U-Net.CT image preprocessing experiments verify the effectiveness of the proposed method.(3)To solve the problem that the traditional fluorescence source reconstruction methods have low reconstruction speed and insufficient accuracy,a deep learning method is applied to fluorescence source reconstruction.According to the data structure characteristics of the three-dimensional fluorescence molecular imaging,the graph convolution is used to construct the fluorescence source reconstruction prediction network.Fluorescence source reconstruction experiments show that the graph convolution network can perform reconstruction quickly and accurately with high robustness.(4)Based on the research results,the small animal optical imaging system is built.The deep learning model proposed in this paper is deployed into the CT image preprocessing and fluorescence source reconstruction module of the system,and the experiments were carried out to verify the performance of the algorithms proposed.Experiments show that the proposed algorithm is more accurate than traditional algorithms.更多还原