基于深度先验学习的头颈一体化磁共振血管壁快速成像关键问题研究

Atherosclerosis plaques occurring in the intracranial and extracranial arteries are the major causes of ischemic stroke. Three-dimensional magnetic resonance vessel wall imaging has been validated to be an effective method for characterizing atherosclerotic plaques. High isotropic spatial resolution, efficient flow suppression, and enough Signal-to-Noise ratio (SNR) are critically required by intracranial and extracranial vessel wall imaging. Moreover, simultaneous imaging of intracranial and carotid arterial walls is clinically preferred since it facilitates the identification of any stroke related plaques in a single scan. However, as the cerebral arteries are very small and are embedded deep in brain tissues, and the signals from both the blood flow inside the vessels and the cerebrospinal fluid may have big influence, whole brain and neck vessel wall imaging is challenging and suffering from the intense tradeoff between spatial coverage, resolution, SNR and scan time, which currently takes 9-10 minutes. Thus fast vessel wall imaging techniques should be developed for this technique to be clinically feasible. This project proposes to achieve fast whole brain and neck vessel wall imaging via deep prior knowledge learning. Precisely, by investigating the interaction mechanism between the learned undersampling trajectory as well as prior knowledge for big datasets and the blood flow, we plan to study the effects of the fusion of the prior knowledge learned from multi-channel and high-dimensional large datasets on the imaging quality. Based on the learned prior knowledge, we are going to develop a high-dimensional reconstruction model with a large field of view. Moreover, we will design the sequence and sampling method of the magnetic resonance imaging accordingly and achieve high-resolution online imaging by considering the multi-dimensional prior knowledge. The outcome of the project is expected to provide technical support for the early prevention and treatment of cerebral ischemic stroke diseases.

颈部和颅内动脉粥样硬化病变是缺血性脑卒中的主要原因，头颈一体化磁共振血管壁成像是针对这些病变的一种重要无创诊断技术。然而，由于脑动脉细小且位于深部脑组织的包埋之中，以及血管内血流信号和脑脊液信号的影响，现有头颈联合血管壁成像技术大多受限于长达9-10分钟的扫描时间，仍无法同时实现头颈一体覆盖、高分辨率、较好的信噪比和对比度。本项目拟针对这一问题，开展基于深度先验学习的头颈一体化磁共振血管壁快速成像研究。具体通过学习欠采样轨迹及大数据先验与血流运动敏感性的相互作用机制，研究多通道大样本高维数据先验的融合对成像质量的影响，开发基于深度先验的大视野高维重建模型，进而设计磁共振成像的序列和采样方式，实现综合多维度先验的在线高分辨成像。项目的实施有望为脑卒中疾病的早期防治提供有力的技术支撑。

颈部和颅内动脉粥样硬化病变是缺血性脑卒中的主要原因，头颈一体化磁共振血管壁成像是针对这些病变的一种重要无创诊断技术。然而，由于脑动脉细小且位于深部脑组织的包埋之中，以及血管内血流信号和脑脊液信号的影响，现有头颈联合血管壁成像技术大多受限于长达9-10分钟的扫描时间，仍无法同时实现头颈一体覆盖、高分辨率、较好的信噪比和对比度。本项目针对这一问题，开展了基于深度先验学习的头颈一体化磁共振血管壁快速成像研究。通过学习欠采样轨迹及大数据先验与血流运动敏感性的相互作用机制，研究多通道大样本高维数据先验的融合对成像质量的影响，开发基于深度先验的大视野高维重建模型，进而设计磁共振成像的序列和采样方式，最终实现综合多维度先验的在线高分辨成像。具体研究内容分为成像和图像分析两部分，在成像方面有：构建了由粗到细的可变形变换无监督多对比度磁共振图像配准框架，设计了基于物理驱动的无监督并行磁共振成像表征学习模型，提出了用于动态磁共振成像的自监督协作学习网络，提出了自监督模型驱动的磁共振图像重建优化算法，开发了基于并行网络训练框架的自监督学习磁共振图像重建方法，构建了基于教师督导的磁共振图像联合重建与分割模型，在较短的扫描时间内获取高分辨率磁共振影像数据。在图像分析方面：设计了一系列用于脑卒中病变分割的模型，构建了一个用于医学图像自动分割的标注高效深度学习框架，设计了基于门控模块的多视图信息融合医学图像分割网络，提出了基于注意力机制的多视图医学图像分类模型，设计了一种可综合利用深度学习和放射组学特征的医学图像分类框架，显著提升磁共振影像数据的分析效果。此外，本项目还构建了一个多通道头颈一体化磁共振样本数据库。本项目的研究为脑卒中疾病的早期防治提供有力的技术支撑，有助于推动相关人工智能技术的进步及临床应用。