多模态婴幼儿脑核磁共振图像海马体分割方法研究

Hippocampus segmentation for multi-modality neonatal brain MR image is one of the most difficult problems in medical image analyze. High accurate volume of the hippocampus and its changes are beneficial to the diagnosis of patients with autism. Due to the effect of severe low contrast, noise, intensity inhomogeneity in the MR images of neonatal, the traditional segmentation model cannot obtain the ideal results. In order to improve the accuracy of multi-modality image information fusion, an image registration and bias field estimation coupled framework, which can preserve the shape feature of tissues, is proposed. A strict concave function, which can be used for multi-modality neonatal brain MR images, is used as the measure function. The structure information is used to improve the low rank representation, which is integrated in to the measure function to preserve the shape characteristic. In view of the characteristics of low rank in the neighborhood of brain MR image, we first research how to construct the structure low rank based deep learning mechanism. Then, we integrate context information into the deep learning mechanism. The local and global characteristic information of the multi-modality neonatal brain MR images can be studied by using the improved deep learning mechanism. In order to deal with the problem of no obvious boundaries between the hippocampus and the adjacent tissues, the deep learning mechanism is used to study the prior information on the shape of the hippocampus. We present an image segmentation and bias field estimation framework, which combines the anisotropic information, characteristic information and shape prior information of the hippocampus, to obtain more accurate segmentation results for hippocampus.

多模态婴幼儿脑MR图像海马体分割是医学图像分割中最困难问题之一。高精度的海马体体积及其变化量有利于辅助诊断自闭症患者患病程度。由于婴幼儿脑MR图像中存在严重低对比度、噪声、偏移场，传统分割模型很难得到理想结果。针对传统配准方法对低对比度、偏移场敏感问题，本项目拟寻找符合多模态婴幼儿脑MR图像特性的严格凹函数作为测度函数并联合具有结构信息的低秩表示以保持目标形状特性，在此基础上构建配准与偏移场矫正耦合框架。针对脑MR图像邻域具有低秩性，本项目拟研究结合结构低秩理论的深度学习机制并在此基础上研究如何将上下文信息融入到深度学习机制中以获取多模态婴幼儿脑MR图像中的局部与整体特征信息。针对海马体与邻近组织无明显边界的问题，拟利用深度学习机制获得海马体形状先验信息。最终在PDE框架下联合各向异性空间信息、局部与整体特征信息以及形状先验信息构建图像分割与偏移场矫正耦合框架以得到高精度海马体分割结果。

项目以婴儿脑MR影像海马体分割作为研究对象，结合有限混合模型、深度学习理论以及水平集理论探讨婴儿脑MR影像分割方法问题。经研究，本项目提出一系列医学影像脑MR影像分割方法，主要包括：（1）提出一套基于限混合模型的脑MR影像分割模型；（2）提出一套基于水平集框架下的脑MR影像分割模型；（3）提出一套基于深度学习框架的脑MR影像分割模型；（4）提出一套基于稀疏表示理论、超像素信息理论等理论的数据分类模型。以上研究表明：重尾偏斜分布族可以更好刻画脑MR影像分布特性；各向异性区域信息可以更好保持局部细节信息；水平集框架下的模型可以得到更为光滑的边界；多尺度深度学习框架可以得到更为准确的婴儿脑MR影像分割结果；低秩表示与超像素结合可以提高分类器精度。实验结果验证了项目前期提出的系统方案有效性。项目资助发表论文共计24篇，其中SCI期刊论文14篇，EI期刊论文2篇，会议论文2篇。培养毕业硕士研究生9名，在读硕士研究生9名。项目投入经费共计75万元，支出42.220194万元，各项支出基本与预算相符，剩余经费21.489926万元，将继续用于项目研究的后续支出。