基于磁敏感加权图像和弥散加权成像的超急性期缺血性脑卒中影像分析研究

Quantification of ischemia from hyper-acute ischemic stroke neuro-imaging is challenging to be explored as it is the prerequisite for patient-specific therapy. Based on theories of image processing, the study is to investigate the segmentation of pathological tissues and fusion of multi-modal images from two imaging modalities which are both sensitive and specific to hyper-acute ischemia, i.e., susceptibility weighted image (SWI) and diffusion weighted imaging (DWI) (including diffusion weighted image and apparent diffusion coefficient map). Starting from what we have achieved on ischemia quantification from DWI, we are going to carry out the following studies. We will learn the grayscale characteristics from deep learning, to add location constraints from building probabilistic atlases, to segment small objects from adding edge constraints, and to aggregate these features and constraints into the framework of a conditional random field, as a new way to solve the long-existing problem of segmenting pathological tissues from images and applying the framework to segment hypointense veins from SWIs. We will employ sparse theory for initial segmentation of ischemia from DWI. We will explore line-shaped profiles to detect hypointense regions from SWI and recognize intracranial thrombus and cerebral microbleeds by virtue of shape and location constraints. We will constrain non-linear registration by adding correspondence between midsagittal planes and maximizing the mutual information in image regions with moderate to significant local grayscale variations. We will fuse the initial ischemia from DWI and veins from SWI to well represent the characteristics of ischemia, and recognize and quantify favorable ischemic penumbra. The study is to provide mathematical and theoretical support for diagnosis and therapy of hyper-acute ischemia based on quantitative imaging features. The theories and methods explored will be valuable for segmenting pathological tissues, nonlinear image registration, and information fusion from multi-modal images.

超急性期脑缺血影像量化困难但却是实现个性化诊疗的关键。本项目基于图像处理理论，利用对脑缺血表征好的磁敏感加权图像SWI与弥散加权成像DWI(包括弥散加权图像和表观弥散系数图)，研究病变组织的分割及多模态图像信息融合。基于我们在DWI量化的研究基础，本项目将通过深度学习获取病变组织的灰度特征、构造概率图谱进行位置约束、添加边缘约束以分割小的目标物、将这些特征约束集成为条件随机场分割框架并分割SWI中的低信号静脉血管；研究基于稀疏理论分割DWI中的初始缺血区域；研究基于直线灰度分布模式检测SWI中的低信号并添加位置与形状约束识别微出血与动脉内血栓；结合中矢状面对应、最大化灰度变化显著区域的互信息实现同一病人脑影像的非线性配准；探索SWI与DWI缺血信息融合并识别、量化良性缺血半暗带。研究成果为建立基于影像量化特征的诊疗提供方法支持，对病变组织的分割、多模态图像的非线性配准及信息融合有参考价值。

超急性期脑缺血量化困难但却是实现个性化诊疗的关键。本项目利用对脑缺血表征好的磁敏感加权图像SWI与弥散加权成像DWI（包括弥散加权图像和对应的表观弥散系数图），研究病变组织的分割及多模态图像信息融合。基于我们在DWI量化的研究基础，提出了结合稀疏表示和包特征的从DWI分割缺血区域的算法将超急性脑缺血的精度由已有的60%提高至75%；提出了一种基于卷积神经网络的磁敏感加权图像SWI静脉分割算法，采用稠密连接技术融合多尺度特征，引入混合目标函数（交叉熵以及Dice损失）解决目标物很小的样本不均衡问题，分割精度达到75.6%；从DWI与SWI两个模态刻画了缺血，提高了超急性期脑缺血量化的灵敏性；研究了基于卷积神经网络的SWI、DWI图像的弹性配准以融合SWI与DWI上面的缺血区域从而提高缺血量化的准确度；提出了融合局部灰度阈值（区域分割）与自适应边缘提取（边界分割）作为初始分割、利用带标记的分水岭方法优化分割的方法，在复杂图像（边界模糊、灰度分布复杂）分割的鲁棒性及精确性方面取得了原创性成果；在深度学习方面，提出了一种融合分类概率与重构误差的分类器；构建了110套超急性脑缺血影像数据库。研究结果为建立基于影像量化特征的超急性期脑缺血的诊疗提供方法支持，丰富了图像分割理论，对病变组织的分割、多模态图像的弹性配准及信息融合有参考价值。