CASIA OpenIR  > 中国科学院分子影像重点实验室
无肿瘤区域引导的生物自发荧光断层成像重建算法研究
高源
Subtype博士
Thesis Advisor田捷
2019-05
Degree Grantor中国科学院大学
Place of Conferral北京
Degree Discipline模式识别与智能系统
Keyword光学分子影像 生物自发荧光断层成像 高斯权重拉普拉斯正则先验 双边权重拉普拉斯正则先验 多层感知机重建模型
Abstract

分子影像技术是一种基于信息科学,生物物理学以及生物医学等多领域理论交叉的计算机成像技术。该影像技术使用靶向性探针或生物医学试剂来标记成像目标的遗传因子、表面受体等生物化学特征,并结合一系列的生物影像获取技术和计算机信号处理方法,进而在组织水平,细胞水平,乃至亚细胞水平上获得生物体内示踪目标的空间分布。因此,分子影像技术能够为活体生物内部的肿瘤位置分布观测,病变组织生长追踪,体内药物代谢以及药效评估等生物医学研究提供更具特异性和灵敏性的成像手段。

生物自发荧光断层成像是一种重要的光学分子影像技术。该断层成像技术以肿瘤等示踪目标的特异性受体与荧光素反应产生的生物荧光作为成像信号。它通过采集由荧光信号传播到生物体表面形成的荧光光斑,进而重建得到荧光光源在生物体内的三维空间分布。传统的生物自发荧光断层成像以光子在生物体内传播过程的数理模型为基础,逆向求解该模态的影像结果。但在实际应用中,现有数理模型对实际光子传播过程的描述存在偏差,且具有严重的欠定性和病态性。这限制了逆向重建的定位精度、形态学信息以及重建鲁棒性等成像性能。此外,基于高分辨率结构模态提供肿瘤先验区域的重建方法虽然能够提升成像结果,却受限于引导模态的影像质量,无法完全体现生物自发荧光的信号特异性。

本文针对如上问题,开展了针对无肿瘤区域引导的生物自发荧光断层成像重建算法的研究。该研究主要分为两个部分:首先,在传统的以描述光子传播过程为基础的重建策略下,参考光源分布规律,设计无肿瘤区域引导的先验正则;其次,在前期研究的基础上,还提出了基于机器学习的重建策略,以数据驱动的方式构建重建模型,进而避免传统策略模型描述存在偏差的缺陷。本文的主要工作和贡献如下:

  1. 提出了基于高斯权重拉普拉斯正则的生物自发荧光断层重建方法,提高了小动物体内荧光光源形态学信息的重建质量。该算法参考弥散性肿瘤的扩散生长方式,假设生物体内空间点之间的荧光光强差异随距离变远而不断增加,进而提出了高斯权重拉普拉斯正则。该先验正则将空间点之间的物理距离转化为高斯权重,并将高斯权重作为径向函数构建非局部拉普拉斯正则矩阵。因此,该矩阵能够对距离较近的空间点对产生较强的荧光光强方差惩罚,而对距离较远的空间点对进行较弱的方差惩罚。相关的数值仿真实验和生物在体重建实验的结果表明,该方法在不依赖其它模态提供的肿瘤区域先验情况下,能够重建得到比传统重建算法更为准确的荧光光源形态学信息。
  2. 提出了基于双边权重拉普拉斯正则的生物自发荧光断层重建方法。该算法的先验正则改良自高斯权重拉普拉斯正则。此正则在生物体内空间点间荧光强度相似性与空间距离呈负相关的原始假设基础上,增加了强荧光区域与弱荧光区域间荧光强度差距较大的先验假设。通过将两个假设相互融合,本算法构建了包含高斯距离权重和高斯强度范围权重的非局部拉普拉斯正则矩阵。该正则矩阵在考虑空间点之间物理距离的同时,还兼顾了荧光的强度变化,进一步提高了强荧光区域与弱荧光区域之间的光强差异。相关数值仿真实验和生物在体重建实验表明,该方法在具有不弱于基于高斯权重拉普拉斯正则重建算法的形态学重建性能的同时,能够得到更为精准的稀疏光源定位结果。
  3. 提出了基于多层感知机的生物自发荧光断层重建方法。不同于传统的基于光学传播模型的重建策略,该算法以多层感知机为基础,借由数据驱动的训练方式,训练得到生物自发荧光断层重建模型。针对传统重建策略中光学传播简化模型的不准确问题,该算法使用蒙特卡洛方法进行更为复杂且准确的光学传播过程模拟,构建训练数据集。针对传统重建策略受限于模型病态性问题,该算法以重建光源与真实光源之间的分布误差作为损失函数,引导重建模型的权重训练,进而避免逆向求解病态的数理模型。此外,为了避免使用蒙特卡洛方法模拟复杂光源样本所需要的庞大计算量,本算法还提出了样本组合方法作为数据扩增手段。该组合方法通过随机组合不同单光源样本的体内荧光光源和表面荧光光斑,以较小的计算消耗构建出大量的多光源样本,为重建模型学习更为复杂的光源情况提供数据支持。通过数值仿真实验与生物在体实验的测验评估可知,相较于传统方法,基于多层感知机的生物自发荧光断层成像算法能够精准定位不同深度,不同间距的荧光光源三维空间位置,具有良好的重建鲁棒性。
Other Abstract

Molecular imaging is a computer imaging technology which based on the intersection of information science, biophysics and biomedicine. It utilizes targeted probes or biomedical reagents to mark the biochemical characterization of imaging target, such as genetic factors and receptors. Combined with a series of biological image acquisition techniques and computer signal processing methods, molecular imaging obtains the spatial distribution of tracer targets in organisms at tissue level, cell level and even subcellular level. Thus, molecular imaging provides more specific and sensitive imaging tools for the biomedical research about tumor location observation, lesion growth tracking, drug metabolism and pharmacodynamic evaluation.

Bioluminescence tomography (BLT) is one of the important optical molecular imaging technologies. It utilizes the bioluminescence produced by the reaction between specific receptors and fluorescein as the imaging signal, and reconstructs the three-dimensional spatial distribution of source in the organism by using the corresponding surface bioluminescence distribution. Conventional bioluminescence tomography reconstruction is an inverse problem of the photon propagation model. However, in practical applications, this model is underdetermined, illness, and has deviations with the true photon propagation. These problems limit the positioning accuracy, morphological recovery and robustness of BLT reconstruction. Furthermore, the reconstruction methods, which based on the tumor region prior referred from the other high-resolution structural modalities, improve the quality of BLT imaging. However, these methods are limited by the imaging quality of these prior, and unable to reflect the signal specificity of bioluminescence completely.

This research focuses on the no tumor region guided BLT reconstruction algorithm to overcome these limitations, and can be divided into two parts: The first part relies on the conventional photon propagation model-based reconstruction strategy. It aims to design a no-tumor-region guided method that only depends on the prior of bioluminescence source distribution. On the basis of no guided method research, the second part proposes a machine learning based reconstruction strategy. This strategy utilizes a data-driven manner to build the reconstruction model, which is used to avoid the defects of model description in conventional strategy. The major contributions of this thesis are listed as follows:

  1. A BLT reconstruction method based on Gaussian weighted Laplace regularization has been proposed to improve the morphological imaging quality of in vivo bioluminescence tomography in small animal. This algorithm depends on the diffusive growth pattern of diffuse tumor, and assumes that the difference of bioluminescence intensity between spatial points in organisms increases as the distance from the increases. Based on this assumption, this algorithm proposes a Gaussian weighted Laplace regularization. This regularization calculates the Gaussian weight depends on the spatial distance between points, and utilizes this weight as a radius function to build a non-local Laplace regularization matrix. Thus, this matrix heavily penalizes the variance of intensity between the closer pairs of points, and lightly penalizes the distant point pairs. The results of numerical simulation experiments and in vivo reconstruction experiments demonstrate that, without the tumor region prior from the other modalities, the proposed algorithm reconstructs more accurate morphological information of bioluminescence source than the conventional algorithms.
  2. A BLT reconstruction method based on bilateral weighted Laplace regularization has been proposed. This regularization is modified from the Gaussian weighted Laplace regularization. Combined with the original assumption that the similarity of bioluminescence intensity between point pair is negatively correlated with their spatial distance, bilateral weighted Laplace regularization assumes a large difference between the points in high intensity region and low intensity region, respectively. Based on this assumption, it builds a non-local Laplace regularization matrix with both Gaussian distance weight and Gaussian range weight. This matrix penalizes the intensity variance not only based on the spatial distance, but also the bioluminescence intensity, and improves the difference between high and low intensity region. According to the evaluation of numerical simulations and in vivo reconstruction experiments, the proposed algorithm obtains more accurate results in sparse source locating, while providing a similar morphological reconstruction performance to Gaussian weighted Laplace regularization method.
  3. A multilayer perceptron-based BLT reconstruction method has been proposed. Different with the conventional reconstruction strategy which depends on the photon propagation modeling, the proposed algorithm utilizes multilayer perceptron (MLP) as the reconstruction model, and trains the model weights in a data-driven manner. To avoid the deviation of model description, this reconstruction strategy collects the training data by applying the Monte Carlo method for simulating the complex but precise photon propagation. To overcome the limitation of illness problem, the proposed method utilizes the loss between reconstructed source and ground truth to guide the network training, rather than calculating the ill-posed inverse problem. Furthermore, MLP based reconstruction method also proposes a sample assemble method as the data expanding method to reduce the computing cost of Monte Carlo simulation. This method builds many cases with complex source by randomly combing the simulated samples with less computational consumption. These assembled data are used to support the reconstruction model in the complex source learning. The results of numerical simulations and in vivo reconstruction demonstrate that, compared with the conventional method, the MLP based BLT algorithm provides more accurate source locating results, and has a robust performance in the reconstruction with different source distribution, such as different source depth, different source gap.
Pages120
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/23811
Collection中国科学院分子影像重点实验室
Recommended Citation
GB/T 7714
高源. 无肿瘤区域引导的生物自发荧光断层成像重建算法研究[D]. 北京. 中国科学院大学,2019.
Files in This Item:
File Name/Size DocType Version Access License
无肿瘤区域引导的生物自发荧光断层成像重建(7003KB)学位论文 开放获取CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[高源]'s Articles
Baidu academic
Similar articles in Baidu academic
[高源]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[高源]'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.