CASIA OpenIR  > 毕业生  > 硕士学位论文
基于卷积神经网络的激发荧光断层成像重建算法研究
黄超
Subtype硕士
Thesis Advisor田捷
2019-05-30
Degree Grantor中国科学院大学
Place of Conferral中科科学院大学
Degree Name工学硕士学位
Degree Discipline计算机应用技术
Keyword激发荧光断层重建, 深度学习, 神经网络, 前向问题, 逆向问题
Abstract

光学分子影像已经成为生物医学研究中一个迅速发展的新兴分子影像技术,
并被广泛的应用于癌症检测、 药物研发和基因表达可视化等方面。 鉴于荧光探针
具有种类多、成本低等特点,荧光分子成像技术已经成为一种很有前景的光学分
子影像技术。然而, 由于荧光探针发射的近红外或可见光光子在经过生物组织时
会被生物组织高度散射,传播到生物体表面所产生的二维荧光图像不能准确的反
应探针的真实分布。因此, 近年来可以重建出荧光探针在生物组织内三维分布的
激发荧光断层成像技术,在理论研究与实际应用中均得以迅速发展。 这种无创、
低成本的实时观测技术,在肿瘤的早期诊断、 病理学、 细胞可视化等领域有巨大
的应用价值。 然而, 扩散方程不能很好的拟合光在非均质生物体内的传输规律以
及重建中逆向问题的不适定性、 病态性都严重限制了激发荧光断层重建技术的重
建效果。 因此,需要针对这两方面的问题做进一步研究。
本论文针对扩散方程不能很好的拟合光在非均质生物组织内的传输规律以
及重建中逆向问题的不适定性、病态性这两方面问题进行研究,充分利用神经网
络强大的函数拟合能力,并把重建中前向问题和逆向问题合并到一个神经网络模
型之中,提出一种新的基于深度学习的激发荧光断层重建算法。该算法直接拟合
生物体表荧光分布与生物体内待重建荧光光源的非线性关系, 通过训练使神经网
络模型自动的拟合光在非均质生物组织内的传输规律,并在训练过程中对样本添
加随机噪声,从而减弱重建中逆向问题的不适定性及病态性。然后使用数值仿真
实验以及在体实验对算法重建性能做了详细的验证。 论文的主要工作包含如下:
1、 传统激发荧光断层重建算法中前向问题、 逆向问题的建立与求解。
光在生物组织内会产生吸收、散射、反射、透射等复杂的相互作用,在实际
应用中一般使用辐射传输方程来近似地描述光在生物组织内的传输规律。 然而,
辐射传输方程的形式十分复杂, 数值求解过程效率低并且占用庞大的计算资源。
因此, 在具体的实际情况下需要对其做出近似和简化。考虑到近红外光在生物组
织内具有高散射、 低吸收的特性, 可以使用计算简单且耗费资源少的扩散方程作
为辐射传输方程的近似模型。鉴于扩散方程属于椭圆型偏微分方程的一种,因此
可以利用有限元法将其转化为矩阵形式的目标方程,然后进行求解。 由于目标方
程具有显著的不适定性以及病态性,因此在逆向问题求解过程中, 需要通过获得
多组生物体表荧光数据、 添加稀疏先验信息或使用贪心求解策略, 来缓解问题的
不适定性以及病态性从而提高重建效果。
2、 基于深度学习的激发荧光断层重建算法研究。
通过对传统激发荧光断层重建算法中前向问题、 逆向问题的研究, 发现除了
逆向问题的不适定性及病态性,扩散方程不能很好的拟合光在非均质生物组织内
的传输规律也会严重限制激发荧光断层重建算法的重建效果。为了提高激发荧光
断层重建算法的重建效果,除了降低逆向问题的不适定性和病态性之外,也有研
究人员使用辐射传输方程的高阶近似模型来描述光在生物组织中的传播规律。虽
然这些方法逐步提高了激发荧光断层重建的重建质量,但近似模型与辐射传输模
型的差异仍会不可避免地导致重建误差。本文充分利用神经网络强大的函数拟合
能力,并把重建中前向问题和逆向问题融合到一个神经网络模型之中,提出了一
种新的基于深度学习的激发荧光断层重建算法, 直接拟合生物体表面荧光数据与
生物体内荧光光源的非线性关系。 通过大量数据对网络模型进行训练, 从而提高
模型对光在非均质生物组织内传输规律的描述精度,并在训练过程中添加随机噪
声, 减弱重建问题的不适定性及病态性。
3、 数值仿真实验及动物在体实验验证。
为了验证算法的可行性, 本文通过数值仿真实验平台构建训练数据集。然后
对深度学习模型进行训练并在训练过程中对样本添加随机噪声。 实验结果表明,
本文提出的方法在荧光光源重心位置重建精度方面优于传统的迭代收缩算法。此
外, 通过改变仿真数字鼠体内器官的分布进一步验证了本文提出的方法对器官分
割时引入的人为误差也具有很好的鲁棒性。为进一步验证算法的实际效果, 本文
设计了在体实验, 并详细地描述了实验设计方案、数据采集和处理方法,并使用
定量指标评价重建结果。 本论文把深度学习用于激发荧光断层重建之中,并且把
前向问题与逆向问题合并为一个模型进行求解, 从而对于如何提高激发荧光断层
重建的重建质量提供了一种新的思路并具有很大的潜力。

Other Abstract

Optical Molecular Imaging (OMI) has become an important and rapidly emerging molecular imaging technology, and has been widely used in cancer detection, drug development and the visualization of gene expression. Based on the variety of fluorescent probes and its low cost, Fluorescent molecular imaging (FMI) has become a promising optical molecular imaging technology. However, due to the high-scattering
and low-absorption of near-infrared or visible-light photons, the two-dimensional FMI cannot accurately reflect true distribution of the probe. Therefore, Fluorescence Molecular Tomography (FMT) which can three-dimensionally reconstruct the distribution of fluorescent probe has been rapidly developed in theory and practical applications. FMT has great application potential and essential research value in the early detection of cancer, pathology research and the visualization of cells. However, the diffusion equations (DE) can not fit the transmission law of photon in heterogeneous biological tissues well and the ill-posedness of the inverse problem limit the reconstruction accuracy of FMT. Therefore, further exploration is needed in these two aspects.
In this thesis, we research on dealing with the problems with the powerful fitting ability of neural network. Besides, we also combined the forward problem and the inverse problem into one neural model. Then a new deep learning based method is proposed to improve the reconstruction efficiency and detailed verifications are performed by using numerical simulation experiments. The main work of the thesis includes the followings:
1. Research on the forward and inverse problems of FMT Light transmitting in biological tissues will undergo complex interactions such as absorption, scattering. In practical applications, RTE is generally used to establish the transmission model. However, the RTE is a complex calculus equation and the solution process is inefficient and requires significant computational resources. Considering the high scattering and low absorption characteristics of near-infrared light in biological
tissues, the diffusion equation (DE) can be used as an approximate model of the RTE. Since the DE is a typical elliptic partial differential equation, it can be transformed into a matrix-form target equation by the finite element method. In order to reduce the significant ill-posedness of the target equation, great amount of effects were made for optimization by adding sparse prior information or using greedy solving strategy.
2. Research on FMT reconstruction algorithm based on deep learning
In order to improve the reconstruction result of the FMT reconstruction, besides reducing the ill-conditionedness of the inverse problem, a high-order simplified spherical harmonics approximation (SPN) can also be used. Although these methods have gradually improved the reconstruction quality of FMT, they still have not achieved
satisfactory results for FMT reconstruction. In the conventional FMT reconstruction, the nonlinear RTE is extensively approximated by DE or SPN, which inevitably causes errors in FMT reconstruction. Therefore, we propose an FMT reconstruction algorithm based on deep learning, which can better reduce the difference between the approximate model and the radiation transfer model.
3. Numerical simulation experiment and in vivo experiments
To evaluate the performance of our proposed method, a 3D digital mouse was utilized to generate FMT Monte Carlo simulation samples. In quantitative analysis, the results demonstrated that our method has better performance than the conventional iterated shrinkage based method in tumor position locating. What’s more, the results
also revealed the good robustness for the varied tumor depth and organ distribution. Besides, in vivo mouse experiments was also carried out to verify the property of our method. To the best of our knowledge, this is the first study that employed deep learning method in FMT reconstruction and also the first study that combined the forward
problem and the inverse problem into one neural model, which holds a great potential of improving the reconstruction quality of FMT.

Pages70
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/23919
Collection毕业生_硕士学位论文
Recommended Citation
GB/T 7714
黄超. 基于卷积神经网络的激发荧光断层成像重建算法研究[D]. 中科科学院大学. 中国科学院大学,2019.
Files in This Item:
File Name/Size DocType Version Access License
毕业论文-黄超.pdf(1947KB)学位论文 暂不开放CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[黄超]'s Articles
Baidu academic
Similar articles in Baidu academic
[黄超]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[黄超]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.