基于多源磁共振成像的高危性斑块研究

Existing intracranial atherosclerotic plaques diagnosis are mainly dependent on the variations in images of the vessel wall based on empirical and qualitative analysis, however such results are usually uncertain. The project intends to study the identification of high-risk intracranial arterial plaque. Combing deep learning algorithms and perceptual organization to resolve the problems such as blur plaque boundaries, overlapping of the main plaque components, and intensity inhomogeneity. The main contents are as follows: build a hierarchical composite morphological structure and integrate the deep learning algorithm to realize the characterization and extraction of plaque morphological features; construct a feature representation of plaque components based on multi-source magnetic resonance imaging data, integrating perceptual grouping rules to study the calculational model to realize the accurate identification of the plaque components; employ the convolution neural network as a technical means to build a prediction model based on plaque load characteristics for high-risk plaque identification.

现有的颅内动脉粥样硬化斑块的诊断主要依赖于医生对管壁影像改变的经验性和定性判断，其结果具有一定的不确定性。本项目拟研究颅内动脉高危斑块的识别，将深度学习、感知组织等结合起来，解决斑块边界模糊、成分重叠和密度值相近等问题。具体包括：构建层次化形状复合结构，融合深度学习算法，实现斑块形态特征的提取与表征；构建多源数据下斑块成分特征表达方式，融合感知组织规则，研究基于感知特征的斑块成分计算模型，实现斑块成分的准确识别；以卷积神经网络为技术手段，构建基于斑块负荷特征的预测模型，实现斑块高危性判别。

1）研究了血管模型的提取方法：提出基于T_frangi增强后血管的连通性，计算连通区域的五元组向量特征，采用浅层神经网络，获取血管模型；此外，考虑到血管模型中孤立点及冗余点的存在，采用改进的vesselness过滤器，量化体素属管状结构的可能性，通过先验形状及局部统计的方法，获取血管树结构。.2）研究了血管狭窄检测方法：采取垂直于中心线血管方向的血管截面序列，提取各截面边缘，标注出可疑狭窄位置并初步判定狭窄等级，之后结合血管单个截面与血管段的狭窄判定结果，对血管狭窄位置及狭窄程度做出进一步判定。此外，从血管中心线角度选取一段血管，计算其邻域所有中心线点坐标并计算主成分，通过切向量进行坐标变换获取血管界面，从而对血管的狭窄程度进行判定。.3）研究了斑块成分及高危性斑块的识别：采用有限元方法计算调和函数及其梯度向量场上的积分曲线，再利用交换图获得血管的内外壁及中间层的表示与呈现获取管壁厚度。同时提取多序列图像的特征，利用统计方法分析不稳定斑块成分的构成以及深度学习方法完成高危性斑块的识别。