中国科学院自动化研究所机构知识库

光学分子影像技术是一种融合了数学、化学、生物学、计算机科学、光学等众多学科的新兴医学成像技术。该成像技术可以实现活体水平下，在组织层面乃至细胞层面对生物的生理变化过程进行动态成像。由于光学分子影像具有空间分辨率高、检测灵敏度高、安全性好等优点，因此自提出之后便取得迅速发展，并被广泛应用于手术导航、药物开发、药效评估等众多领域。
激发荧光断层成像技术是光学分子影像技术的重要组成之一。该成像技术利用荧光探针对肿瘤特定靶标分子标记作用进行体表荧光成像，再使用重建算法结合断层图像进行三维重建，最终实现体内荧光探针分布的三维动态观测。相较于激发荧光成像技术，该成像技术通过融合断层成像信息来提高激发荧光成像技术的深度重建精度。传统的激发荧光断层重建方法依赖于光子在生物体内传播的数学模型，主要通过对辐射传输方程使用一阶球谐简化，来获取光子传播的系统矩阵，再根据系统矩阵与表面荧光强度分布向量求解生物体内光源。然而通过球谐近似求取系统矩阵会存在一定的误差，最终会影响激发荧光断层重建的精度。此外，传统重建方法主要使用基于正则化先验的坐标下降方法进行求解，求解结果普遍出现重建过稀疏或过平滑、空间不连续、形态学重建精度低等问题。因此，本文针对以上问题，开展了基于图网络的脑胶质瘤激发荧光断层重建算法研究。该算法通过数据驱动的方式来训练重建模型，避免了构建光子传输模型和求解逆向问题时出现精度下降的问题。同时，根据脑胶质瘤在生物体中往往成簇成团分布的空间结构先验信息，使用图网络算法提取生物体的空间结构信息来优化激发荧光断层重建的最终重建结果，以达到从定位精度、形态学误差、鲁棒性等多个方面优化重建算法性能的目的。本文的主要研究内容和创新贡献归纳如下：
1、提出一种基于图卷积残差连接网络的激发荧光断层重建算法
在癌症病人体内，肿瘤的生长往往呈现成团状分布的特性，因此，靶向特定肿瘤的荧光探针在病人体内也存在成团成簇分布的结构先验信息。基于该先验信息，本文提出基于图卷积残差连接网络的激发荧光断层重建算法。该算法的主要创新点在于其使用全连接网络拟合体表荧光分布与体内荧光探针分布的非线性映射关系进行体内肿瘤分布的初步定位，并使用图卷积网络融合图中相邻节点特征的特性对初步定位结果进行形态学优化，提高整体形态学重建精度。为验证该算法的重建性能，本文设计了测试集单光源样本以及大光源样本定量实验、脑胶质瘤数字小鼠双光源重建实验和真实脑胶质瘤小鼠在体实验。实验结果表明，该算法在反问题模拟算法基础上显著提高了三维重建的形态学重建精度。
2、提出一种基于Graph-Unet残差连接网络的激发荧光断层重建算法
在传统激发荧光断层重建算法中对断层结构信息的利用仅局限于使用其进行组织剖分以及四面体网格生成，未利用其结构信息来提高激发荧光断层重建精度。为充分利用解剖结构信息，本文提出基于Graph-Unet残差连接网络的激发荧光断层重建算法。该算法的主要创新点在于其利用Graph-Unet的encoder-decoder结构提取不同尺度下的解剖结构信息，并利用基于注意力机制的图注意力网络来融合体表荧光特征和解剖结构特征来提高激发荧光断层重建的精度。针对该算法，本文基于脑胶质瘤数字小鼠设计了不同噪声下单光源重建实验、不同间距双光源重建实验、不同深度复杂大光源重建实验，同时还进行真实脑胶质瘤小鼠在体实验。实验结果表明，本方法相较于对比方法，不仅可以取得准确的定位精度，还能有效地缓解重建过稀疏以及过平滑所导致的形态学重建性能下降的问题，提高其三维形态学重建精度。

Optical molecular imaging is an emerging medical imaging technology that integrates mathematics, chemistry, biology, computer science, optics and many other disciplines. This imaging technology enables in vivo level dynamic imaging of physiological processes at the tissue level and even at the cellular level. Due to the advantages of high spatial resolution, high detection sensitivity and good safety, optical molecular imaging has been developed rapidly since its introduction and is widely used in many fields such as surgical navigation, drug development and drug efficacy assessment.
Fluorescence molecular tomography (FMT) is an important imaging modality in optical molecular imaging techniques. This imaging technique uses fluorescence probes for surface fluorescence imaging of tumour-specific target molecular markers, and then uses reconstruction algorithms combined with tomography for three-dimensional reconstruction to achieve a three-dimensional dynamic observation of fluorescence probe distribution in vivo. This imaging technique improves the reconstruction accuracy of FMT by incorporating tomographic techniques compared to two- dimensional fluorescence molecular imaging. Traditional FMT reconstruction methods rely on a mathematical model of photon propagation in an organism, mainly through a first-order spherical harmonic simplification of the radiative transfer equation to obtain the system matrix of photon propagation, and then solve for the in-organism light source based on the system matrix and the surface fluorescence photon density distribution vector. However, there is error in obtaining the system matrix through the spherical harmonic approximation, which ultimately affects the accuracy of the FMT reconstruction. In addition, traditional reconstruction methods mainly use coordinate descent methods based on a priori of regularization to solve, and the results generally show problems as over-sparse or over-smooth, spatial discontinuity and low accuracy of morphological reconstruction. Therefore, this work addresses the above problems and conducts a study of graph network-based algorithms for FMT reconstruction of glioblastoma. At the same time, based on the priori information of the spatial structure of glioblastoma, which are often distributed in clusters in organisms, the graph network algorithm is used to extract the spatial structure information of organisms to optimize the final reconstruction results of FMT reconstruction with hig localization accuracy, low morphological error, improved robustness and other aspects. The aim is to optimize the performance of the reconstruction algorithm in terms of localization accuracy, morphological error, robustness, etc. The main research content and innovative contributions of this work are summarised as follows.
1. FMT reconstruction algorithm based on Graph Convolutional residual connection network
In cancer patients, the tumours tends to show a cluster-like distribution, therefore, fluorescence probes targeting specific tumours also have a priori information about the structure of cluster-like distribution in patients. Based on this priori information, this work proposes a FMT reconstruction algorithm based on a graph convolutional network residual connection network. An important innovation of the algorithm is its use of a fully connected network to fit the nonlinear mapping relationship between the distribution of surface fluorescence and the distribution of fluorescence probes in vivo. In addition, it also uses a graph convolutional network to fuse the characteristics of adjacent nodes in the graph to morphologically optimize the preliminary localization results and improve the overall morphological reconstruction accuracy. To verify the reconstruction performance of the algorithm, single-source as well as large-source sample quantitative experiments, dual-light reconstruction experiments of digital mice and real glioma bearing mice experiments were designed for the test set. The experimental results show that the algorithm significantly improves the morphological reconstruction accuracy of 3D reconstruction based on the IPS algorithm.
2. FMT reconstruction algorithm based on the Graph-Unet residual connectivity network
In the traditional FMT reconstruction algorithm, the use of tomographic structure information is only limited to tissue dissection and tetrahedral mesh generation, but the structural information is not used to improve the accuracy of FMT reconstruction. In order to make full use of the anatomical structure information, this work proposes a FMT reconstruction algorithm based on Graph-Unet residual connectivity network. The encoder-decoder structure of Graph-Unet is used to extract the anatomical structure information at different scales, and the attention-based graph attention network is used to fuse the surface fluorescence features with the anatomical structure features to improve the accuracy of the FMT reconstruction. For this algorithm, single-source reconstruction experiments with different noise levels, double-source reconstruction experiments with different spacing and complex large-source reconstruction experiments with different depths were designed based on glioma digital mice, and real glioma mice in vivo experiments were also conducted. The experimental results show that this method can effectively alleviate the degradation of morphological reconstruction performance caused by over-sparse and over-smooth reconstruction compared with the comparative method, and thus improve the accuracy of its 3D morphological reconstruction.