AI enabled motion corrected quantitative MRI of the fetal brain

博士项目的目的：在怀孕的后半段，人类的大脑经历了旺盛的生长，微观和宏观的变化迅速发生。因此，作为关键组织特性的MRI弛豫时间（T1和T2）发生了显著变化。这些弛豫时间可用于抑制显影，特别是在白色物质中。弛豫测量法是用于评估健康和疾病中的大脑的成熟工具，然而，目前还没有在子宫内的胎儿中这样做的成熟方法。该项目的目的是发展可靠的运动宽容的方法来测量T1和T2在移动的胎儿，并部署这些进行系统的定量研究，在胎龄，从而提供规范的基准data.Project描述/背景：我们的小组有大量的跟踪记录，在开发和部署完全运动校正的方法胎儿三维脑成像在子宫内。一种非常有效的策略是对单个切片或一小组切片进行快照成像，获取的速度足以冻结胎儿运动，然后使用切片到体积重建（SVR）1重新对齐以校正头部位置的变化。过去的研究包括开发用于解剖、扩散和功能成像的综合方法。我们还能够通过将标准物理信号模型直接拟合到使用专用多回波方法采集的单次激发切片来测量弛豫时间T2*。该方法允许将模型拟合与运动校正分离。已经为心脏（子宫外）开发了用于测量运动组织中的T1和T2的方法，但是这些方法通常依赖于心脏运动的受约束的重复性质。由于胎儿运动的不可预测性，测量胎儿大脑中的这些参数构成了一个特殊的挑战。这些参数也比T2* 更难测量，因为拟合适当的弛豫模型可能需要组合在多次发射中获得的具有不同对比度的图像，并且控制自旋物理学更复杂。同样重要的是，所开发的任何方法都具有低RF功率沉积，控制母亲外周神经刺激的风险，并且是时间有效的，因为延长的胎儿检查对于孕妇来说可能是具有挑战性的，并且随着更全面的MRI方法的开发，这些检查的范围正在日益扩大。因此，该项目将涉及研究新的序列，以实现优化的采集策略，并开发重建方法，允许联合估计运动参数和所需的松弛参数。过去的重建方法对计算要求非常高，因此速度很慢，因此我们建议探索机器学习方法，将计算负担从检查时间转移到训练阶段，从而实现临床上更可接受的快速图像生成。这将建立在我们之前基于深度学习的电影心脏图像重建工作的基础上，该工作提供了最先进的性能，并且最近开始将运动校正作为重建的一部分。我们还探索了深度学习方法在解剖成像SVR中的应用，因此为构建当前项目奠定了坚实的基础。如果能够以足够高的分辨率实现稳健的定量T1和T2映射，那么通过利用血红蛋白的特定弛豫特性来映射氧提取将变得可行。胎儿血管造影术的最新结果提供了令人鼓舞的证据，它是可行的，以解决胎儿血管，虽然实现这与定量方法将是一个拉伸目标。

Aim of the PhD Project:During the second half of pregnancy the human brain undergoes exuberant growth, with both microscopic and macroscopic changes happening rapidly. In consequence the MRI relaxation times (T1 and T2), which are key tissue properties, change substantially. These relaxation times can be used to characterise development, particularly in white matter. Relaxometry is a well established tool for assessing the brain in health and disease, however, there are currently no established methods for doing this in the fetus in utero. The aim of the project is to develop reliable motion tolerant methods for measuring T1 and T2 in the moving fetus, and to deploy these to conduct systematic quantitative studies over gestational age and so provide normative benchmark data.Project Description / Background:Our group has substantial track record in developing and deploying fully motion corrected methods for fetal 3D brain imaging in utero. A highly effective strategy is snapshot imaging of individual slices or small groups of slices, acquired fast enough to freeze fetal motion, which are then realigned to correct for changes in head position using slice to volume reconstruction (SVR)1. Past research has included developing comprehensive methods for anatomical, diffusion and functional imaging. We have also been able to measure the relaxation time T2* by fitting a standard physics signal model directly to single shot slices acquired using a dedicated multi-echo methodology. This approach allows the model fiting to be separated from motion correction. Methods for measuring T1 and T2 in moving tissue have been developed for the heart (ex utero) but these generally rely the constrained repetitive nature of cardiac motion. Measuring these parameters in fetal brain constitutes a special challenge because of the unpredictable nature of fetal motion. These parameters are also harder to measure than T2* as fitting the appropriate relaxometry models is likely to require combination of images with different contrasts acquired over multiple shots and controlling the spin physics is more complex. It is also critical that any methods developed have low RF power deposition, control risk of peripheral nerve stimulation in the mother and are time efficient as prolonged fetal examinations can be challenging for pregnant mothers, and increasingly the scope of these examinations is expanding as more comprehensive MRI methods are developed. The project will thus involve both research into novel sequences to achieve optimised acquisition strategies and also development of reconstruction methods that allow joint estimation of motion parameters and the desired relaxation parameters. Past reconstruction methods have been extremely computationally demanding and hence slow, so we propose to explore machine learning methods to shift the computational burden from examination time to a training phase allowing a much more clinically acceptable rapid image generation. This will build on our prior work on Deep Learning based reconstruction of cine cardiac images, which delivered state of the art performance, and has recently started to include motion correction as part of the reconstruction. We have also explored the application of Deep Leaning methods to SVR for anatomical imaging, so have a strong basis from which to build the current project. If robust quantitative T1 and T2 mapping can be achieved at high enough resolution, then it will become feasible to map oxygen extraction by exploiting the specific relaxation properties of haemaglobin. Recent results on fetal angiography provide encouraging evidence that it is feasible to resolve fetal vessels, although achieving this with quantitative methods will be a stretch target.