多源影像中脑结构的时空一致表示与学习模型研究

With the development of imaging technology, medical imaging data can be acquired from multiple sources. For example, brain images can be acquired from complementary imaging modalities at varying time points. In this context, it is potentially valuable to conduct research on analyzing multi-source brain images for the following reasons. On one hand, it is difficult to obtain accurate semantic segmentations of brain images with a single source due to the incomplete information of target structures. On the other hand, images from multiple sources naturally contain complementary visual patterns of the target, which paves the way to exploring their correlations and developing a novel unified representation model. To this end, we aim to conduct research on unified spatial-temporal representation for brain structures in multi-source images, where the key issue is how to effectively combine the multi-source images and uniformly represent them. Specifically, we first plan to build the spatial-temporal correlation model between multi-modality features based on the multi-modal longitudinal brain imaging data. Second, we aim to learn and extract the common discriminative features from multi-source images, which are used in the task of accurate semantic segmentation of brain structures. Finally, we will validate the effectiveness of the developed models in applications, such as computer-aided analysis. Our research is expected to provide theoretical and technical support to the image-based analysis for early brain development of infants.

随着成像技术不断发展，医学影像数据呈现显著的多源特性，如多模态、多成像时间点等。这一特性为脑影像分析研究带来了机遇：单源图像因其携带信息有限，很难实现待分析目标的高质量语义提取；而多源影像为目标提供了互补性的多种视觉呈现，充分挖掘它们之间的关联，应能形成更为精确的目标表示。而如何对多源影像进行协同分析和统一表示是其中的关键所在。针对这一问题，项目拟研究多源影像中脑结构的时空统一表示与学习模型。具体包括：以多时间点多模态的脑影像为研究对象，通过时空一致性约束构建多模态特征间的高维关联模型；研究多源影像中高判别力特征的学习与提取，以实现脑结构的精确语义分割，并在相关计算机辅助分析等应用中加以验证。研究结果有望为新生儿脑部早期发育过程的影像分析提供理论基础和关键技术支持。

医学影像数据具有显著的多源影像特性，如多分辨率、多模态、多采集时间点等，为医学图像研究带来了机遇和挑战。本项目以多源脑影像数据为研究对象，探索多源影像中目标统一表示机制，构建对数据集鲁棒并且精确的图像表示建模与学习方法，能够在充满类内/类间变化和成像干扰的复杂模式中精确突显出多模态特征的关联和互补。在多源影像特征表示与关联建模、多模态特征的语义分割、高判别力特征的学习与提取以及图像增强模型构建等方面取得了一定的研究成果。提出多图谱下空间超图构建，实现半监督超图学习的语义分割，应用于成人和婴幼儿脑部组织结构提取；提出利用Unet将高层特征和浅层特征在深度卷积神经网络中进行多通道的底层融合与多特征的高层语义融合，实现对复杂表征下特征之间的语义关联进行挖掘，应用于脑部海马体组织分割之中；提出自适应稀疏结构下的特征学习，实现在统一框架下同时学习特征矩阵和数据流形结构，应用于神经精神疾病辅助诊断；提出采用半解耦的方式完成高质量的Retinex分解，实现基于Retinex图像表示的暗光增强。上述研究成果发表于领域内知名国际期刊共十篇。