Highly Accelerated Magnetic Resonance Angiography using Deep Learning

血管成像（血管造影）在脑部尤为重要，因为脑部的血液供应紊乱和出血会造成严重后果。血管造影可以同时显示血液供应中断（动脉粥样硬化、栓塞）和血管异常（动脉瘤、动静脉畸形）。然而，传统的血管造影技术需要注射造影剂并暴露于电离辐射中，这对患者有一定的风险。某些基于磁共振成像（MRI）的方法没有这些缺点：例如，飞行时间（TOF）和动脉自旋标记（ASL）血管造影术。ASL不如TOF那么成熟，但可以提高血管的可视性，并允许获得动态的、血管选择性的血管造影。然而，这两种方法的扫描时间都很长，特别是当需要高空间分辨率和全头部覆盖时，这使得它们难以适应繁忙的临床协议，并且由于患者的运动增加了图像损坏的可能性。在这个项目中，将开发高度加速的血管造影方法，将原始信号（k空间）域的欠采样与基于物理信息监督深度学习的新型图像重建方法相结合。我们预计，大脑中非常特定的血管分支结构可以通过深度卷积网络很好地表示，从而实现高空间分辨率和全头部覆盖的快速扫描，而不会产生传统（并行成像或压缩感知）欠采样重建相关的伪影和长时间处理时间。这将包括优化展开迭代网络架构，结合欠采样模式，对大型回顾性欠采样TOF数据集进行培训和测试，对ASL血管造影进行微调，并将其应用于健康志愿者和患者队列的前瞻性欠采样快速扫描数据。该框架也将适用于血管识别，允许对获得的血管造影数据（例如分支模式，血管弯曲）进行定量分析，并为基于mri的动脉硬度脉冲波速度测量提供关键信息，这被认为与血管性痴呆有关。

Imaging of the blood vessels (angiography) is particularly important in the brain, where disturbances in blood supply and haemorrhage have severe consequences. Angiograms allow the visualisation of both blood supply disruption (atherosclerosis, embolism) and vascular abnormalities (aneurysms, arteriovenous malformations). However, conventional angiographic techniques require the injection of a contrast agent and exposure to ionising radiation, resulting in some risks to the patient. Certain magnetic resonance imaging (MRI)-based methods do not have these drawbacks: for example, time-of-flight (TOF) and arterial spin labelling (ASL) angiography. ASL is less well established than TOF, but improves vessel visibility and allows dynamic, vessel-selective angiograms to be obtained. However, both suffer from long scan times, particularly when high spatial resolution and whole-head coverage are required, making them difficult to fit into busy clinical protocols and increasing the likelihood of image corruption due to patient motion.In this project highly accelerated angiographic methods will be developed that combine undersampling in the raw signal (k-space) domain with novel image reconstruction methods based on physics-informed supervised deep learning. We anticipate that the very specific branching structure of vessels within the brain can be well represented by deep convolutional networks, allowing rapid scans with high spatial resolution and whole-head coverage without the artefacts and long processing times associated with conventional (parallel imaging or compressed sensing) undersampled reconstructions.This will involve optimising unrolled iterative network architectures in combination with undersampling patterns, training and testing on large retrospectively undersampled TOF datasets, fine-tuning for ASL angiography, and application to prospectively undersampled rapid scan data in healthy volunteers and patient cohorts. This framework will also be adapted for vessel identification, allowing quantitative analysis of the acquired angiographic data (e.g. branching patterns, vessel tortuosity) and providing crucial information for complementary MRI-based pulse wave velocity measurements of arterial stiffness, thought to be implicated in vascular dementia.