中国科学院分子影像重点实验室知识库

      光学分子影像技术利用分子探针或者生物蛋白对生物体进行标记。在特定条件下，生物体内能够释放出荧光信号，光学分子影像技术通过采集这种荧光信号来获取生物体内分子细胞水平的生理信息。光学分子影像技术具有高灵敏度、高特异性、无辐射、价格低廉等优势。目前，该技术已被广泛应用于肿瘤检测、药物研发、手术导航等临床和预临床研究当中。

      生物自发荧光断层成像 (Bioluminescence tomography, BLT) 技术是光学分子影像中的一项重要内容。相比于二维的自发荧光成像技术而言，该技术可以获得成像对象的三维空间分布信息，进而为研究人员提供更为丰富的影像信息。然而，由于光在生物组织传播过程中会发生强烈的散射作用，且生物体表的测量数据十分有限，这些问题导致生物自发荧光断层成像是一个严重的病态性问题。因此，BLT重建的准确性、稳定性和实用性等方面仍然面临着很大挑战。在实际中，往往需要引入光源空间分布的先验假设来缓解病态性问题，其中稀疏先验被广泛应用于BLT重建领域，然而该先验信息引导重建结果呈稀疏分布，虽然能获得准确的光源位置信息，但是丢失了大部分的光源形态信息，这限制了BLT在实际追踪空间特异性分布任务中的应用价值。本文基于生物自发荧光断层成像光源的空间分布特性，在块稀疏贝叶斯框架下开展了一系列重建算法研究。本文的主要工作和创新点如下：

1. 小动物多模态成像系统及在体实验基本流程。小动物在体实验是生物自发荧光断层成像研究中的一项重要内容，也是验证成像方法的检验标准。本文详细介绍了由本团队自主研发的小动物多模态成像系统及基于该系统所开展的一般在体实验的基本流程，其中包括小动物模型建立过程，以及基于相关成像系统完成的数据采集和数据处理流程。

2. 基于K均值聚类策略的块稀疏贝叶斯重建算法。该算法在引入重建光源稀疏先验的基础上，进一步引入光源呈聚类分布的结构先验假设。基于该假设，首先使用K均值聚类策略直接将空间三维网格节点根据空间距离进行聚类分组，构建基于K均值聚类的块稀疏先验模型。同时，本研究考虑到网格节点之间的相关性与空间距离呈反比关系，即假设空间距离越小的节点之间空间相关性越强，因此，将高斯加权距离先验引入到块内相关性矩阵的设计中，用以描述块内邻居之间 (邻域内) 的相关性。相关的数字鼠仿真实验和小鼠原位脑胶质瘤实验结果表明，该方法在保证光源定位精度的同时，缓解了传统稀疏算法的过收缩特性，能够恢复出更多的光源分布信息，进一步提高了BLT形态学重建精度。

3. 基于K近邻分组策略的块稀疏贝叶斯重建算法。该算法同样引入了光源稀疏、聚类分布假设以及网格节点的空间邻域相关性。不同于直接对空间网格节点进行粗聚类，该算法采用一种更为细化的分组策略，即将每一个网格节点看做一个聚类中心，基于K近邻算法，选择距离聚类中心最近的K个邻居节点，由此构建基于K近邻策略的块稀疏先验模型。由于该算法不涉及任何随机过程，因此相比于基于K均值聚类的块稀疏贝叶斯重建算法而言，能保证算法的稳定性。为了提高重建精度，该算法设置了更为精细的聚类任务，可以进一步保留光源信息，避免大多数光源能量值在算法迭代中被置为零。相关数字鼠仿真实验和小鼠原位脑胶质瘤实验表明，该方法在保证重建结果稳定性的同时，能够实现准确的形态学重建。

4. 基于自适应分组策略的块稀疏贝叶斯重建算法。该算法是对基于K近邻策略的块稀疏贝叶斯重建算法的改进算法。在引入光源稀疏、聚类分布假设及网格节点邻域相关性的基础上，引入迭代过程中节点荧光能量值的信息。该算法仍是将每一个网格节点看作一个聚类中心，然后基于自适应分组策略，在算法迭代过程中不断调整分组状态。该自适应策略基于两个准则来实现：准则1：将与目标节点共享同一个四面体的所有节点作为初始邻居节点，由此提供一个初始的粗分组状态。准则2：基于设置的能量阈值，在迭代过程中，对组内邻居进行二次约束，即剔除掉能量值小于阈值的邻居节点。该自适应分组模型的优势在于，首先根据准则1定义一个宽泛的解空间，准则2进一步对解空间进行细化，去掉一些无效的邻居节点，在算法迭代过程中，不断优化求解空间。该算法不用事先确定任何分组参数，并能根据不同个体进行自动分组及调整。相关数字鼠仿真实验，仿体实验和小鼠在体实验表明，该方法在保证定位精度和形态学重建精度的基础上，进一步提升了算法的稳定性、灵活性和实用性。

    Optical molecular imaging technology uses molecular probes or biological proteins to label organisms. Under certain conditions, these organisms can release bioluminescent signals. Optical molecular imaging technology collects these bioluminescent signals to obtain physiological information at the molecular and cellular level in organisms. Optical molecular imaging technology has the advantages of high sensitivity, high specificity, no radiation, and low price. At present, this technology has been widely used in clinical and pre-clinical studies such as tumor detection, drug development, and surgical navigation.

    Bioluminescence tomography (BLT) is an important content in optical molecular imaging. Compared with the two-dimensional bioluminescent imaging technology, this technology can obtain the three-dimensional spatial distribution information of the imaging object, thereby providing researchers with richer image information. However, due to the strong scattering of light during the propagation of biological tissues, and the limited measurement data on the surface, BLT is a serious ill-posed problem. The accuracy, stability, and practicality of the reconstruction results are still facing great challenges. In practice, it is often necessary to introduce a priori assumptions about the spatial distribution of light sources to alleviate the ill-posed problem. Sparse prior is widely used in the field of BLT reconstruction. However, the prior information guides the reconstruction results to be sparsely distributed, although accurate position information can be obtained, most of the light source morphology information is lost, which limits the application values of BLT in the tracking spatial-specific distribution tasks. In this thesis, based on the distribution characteristics of the light source for BLT, a series of reconstruction algorithms are proposed under the block sparse Bayesian framework. The main work and innovations of this thesis are as follows:

1. Small animal multi-modality imaging system and basic flow of in vivo experiments. In vivo experiments on small animals are an important part of the study of BLT, and it is also a test standard for verifying imaging methods. This thesis introduces in detail the small animal multi-modality imaging system independently developed by our team and the basic process of general in vivo experiments carried out based on the system. This includes the process of establishing a small animal model, as well as the data acquisition and data processing process based on the relevant imaging system.

2.Block sparse Bayesian reconstruction algorithm based on K-means clustering strategy. Based on the introduction of the sparse prior of the reconstructed light source, the algorithm further introduces the structural characteristics of the light source in a clustered distribution. Based on this assumption, we first use the K-means clustering strategy to directly cluster the spatial three-dimensional mesh nodes according to the spatial distance and build a block sparse prior model based on K-means clustering. At the same time, we consider that the correlation between mesh nodes is inversely proportional to the spatial distance, that is, it is assumed that the smaller the spatial distance, the stronger the spatial correlation between nodes. Therefore, the Gaussian weighted distance prior is introduced into the intra-block correlation matrix, it is used to describe the correlation between the neighbors in the block. We conducted related digital mouse simulations and mouse orthotopic glioma experiments. The experimental results showed that this method can not only ensure the positioning accuracy of the light source, but also alleviate the over-sparseness of the traditional sparse algorithm, and then restore more source distribution information further improve the morphological accuracy of the reconstruction. 

3. Block sparse Bayesian reconstruction algorithm based on K-nearest neighbor clustering strategy. The algorithm also introduces the assumptions of light source sparseness, clustering distribution, and spatial neighborhood correlation of mesh nodes. Different from the rough grouping of spatial mesh nodes directly, this algorithm adopts a more refined grouping strategy, that is, regarding each mesh node as a clustering center, based on the K-nearest neighbor algorithm, the K neighbor nodes closest to the cluster center are selected, thereby constructing a block sparse prior model based on the K-nearest neighbor strategy. Since the algorithm does not involve any random process, compared with the block sparse Bayesian reconstruction algorithm based on K-means clustering, it can ensure the stability of the algorithm. In order to improve the reconstruction accuracy, we set up a more refined clustering task, which can further retain the light source information, avoiding most of the light source intensity value being zero in the algorithm iteration. Related digital mouse simulations and mouse orthotopic glioma experiments showed that this method can further achieve morphological reconstruction while ensuring the stability of the reconstruction results. 

4. Block sparse Bayesian reconstruction algorithm based on adaptive grouping strategy. This algorithm is an improved algorithm for the block sparse Bayesian reconstruction algorithm based on the K-nearest neighbor strategy. On the basis of the introduction of light source sparseness, clustering distribution characteristics, and mesh nodes correlation, this algorithm also introduces nodes intensity value in the iterative process. This algorithm still regards each mesh node as a clustering center, and then continuously adjusts the grouping state during the algorithm iteration process. The adaptive strategy is implemented based on two criteria: Criteria 1: All nodes that share the same tetrahedron with the target node are regarded as initial neighbor nodes, thereby providing an initial coarse grouping state. Criterion 2: Based on the intensity value threshold, in the iterative process, the neighbors in the group are constrained twice, that is, neighbor nodes whose intensity value is less than the threshold are eliminated. The advantage of this adaptive grouping model is that it first defines a broad solution space according to criteria 1, and criteria 2 further refines the solution space, removing some invalid neighbor nodes, and continuously optimizing the solution space during the solution process. The algorithm does not need to determine any grouping parameters in advance. During the entire algorithm process, it can automatically group and adjust flexibly according to different individuals. Related digital mouse simulations, phantom experiments, and in vivo mouse experiments showed that this method further improved the stability, flexibility, and practicability while ensuring positioning accuracy and morphological reconstruction accuracy.