Development of a fully automatic, 20-second long deep-learning based calibration procedure for parallel transmission (pTx) in ultrahigh field MR body imaging

在临床磁共振成像（MRI）中，通常使用磁场强度为1.5和3特斯拉的MRI扫描仪。此外，所谓的超高场（UHF） MRI扫描仪工作在7特斯拉及以上，越来越多地被研究用于临床和科学用途，其中包括更高的空间图像分辨率和更快的图像采集。虽然7特斯拉的优势已经被用于头部或四肢的临床诊断应用，但7特斯拉对身体成像的好处却很少被研究，而且还没有被批准用于常规临床检查。造成这种情况的主要原因和超高频身体MRI的一个主要问题是图像的空间高度不均匀的信号。这种效应是由射频（RF）场的磁分量（B1+）的不均匀分布引起的，射频（RF）场被射频天线照射以成像核自旋。虽然这种不均匀分布可以通过所谓的“并行传输（pTx）”来非常成功地补偿，即通过使用多个天线并通过单独的RF脉冲独立驱动它们，但这需要两个校准步骤：在拍摄快速概述图像（定位器）之后，对于每个患者，必须测量每个天线的B1+图，ii)必须计算pTx RF脉冲。这种方法的缺点是校准时间长，通常在体内需要10-15分钟，其中大部分时间都在步骤1)。正是这段校准时间严重阻碍或阻碍了科学研究，也阻碍了超高频在体内的临床应用。大量减少这个时间是这个应用程序的目标。拨款申请是基于申请人小组最近提出的一项初步技术，其中B1+地图不需要通过单独扫描进行测量。相反，通过使用神经网络（NN），从定位器地图中估计地图，无论如何，这是在研究开始时获得的。在此应用中，将进一步发展该技术，并将系统地分析其精度和鲁棒性。此外，在与奥胡斯大学的合作进行的单独工作中，pTx RF脉冲计算也使用神经网络进行了大规模加速。在目前的拨款申请中，后一种技术将与先进的基于神经网络的B1+映射方法相结合，形成一个单一的基于人工智能的校准步骤，校准时间不超过20秒。这种新方法将在柏林、海德堡和美国明尼阿波利斯的不同超高频中心进行测试。最终的校准方法允许首次进行针对人体的UHF患者研究，而不需要长时间的校准，这将促进未来的患者研究，以调查7特斯拉对人体的益处。

In clinical magnetic resonance imaging (MRI), MRI scanners with a field strength of 1.5 and 3 Tesla are typically used. In addition, so-called ultra-high field (UHF) MRI scanners operating at 7 Tesla and beyond are increasingly being investigated for clinical use as well as for scientific purposes, which allow, amongst others, higher spatial image resolution and faster image acquisition. While the advantage of 7 Tesla is already used diagnostically for clinical applications in the head or extremities, the benefits of imaging the body at 7 Tesla has been investigated rather rarely and it has not yet been approved for routine clinical examinations. The main reason for this and a major problem in UHF body MRI is the spatially highly inhomogeneous signal of the image. This effect arises from the inhomogeneous distribution of the magnetic component (B1+) of the radiofrequency (RF) fields irradiated by the RF antenna to image the nuclear spins. Although this inhomogeneous distribution can be compensated very successfully by so-called "parallel transmission (pTx)" i.e. by using multiple antennas and driving them independently by separate RF pulses, this requires two calibration steps: after taking a fast overview image (localizer), for each patient i) the B1+ map for each antenna has to be measured and ii) the pTx RF pulses have to be calculated. A drawback of this approach is the long calibration time of typically 10-15 minutes in the body, most of which falls on step i). It is this calibration time that severely hinders or prevents not only scientific studies but also clinical UHF applications in the body. Massively reducing this time is the goal of this application. The grant application is based on a recently by the group of the applicant presented preliminary technique, in which the B1+ maps do not need to be measured by a separate scan. Instead, by using neural networks (NN) the maps are estimated from the localizer maps, that is anyway acquired at the beginning of the study. In this application, this technique will be developed further and will be systematically analyzed with respect to precision and robustness. Furthermore, in a separate work performed jointly with collages from Aarhus University, the pTx RF pulse calculation has been massively accelerated using NN, too. In the present grant application, the latter technique will be combined with the advanced NN-based B1+ mapping method to form a single AI-based calibration step that does not require more than 20 seconds of calibration time. This novel method will be tested at different UHF centers in Berlin, in Heidelberg and in Minneaplis, USA. The final calibration method allows for the first time to perform UHF patient studies targeting the human body without the need of lengthy calibration, which will promote future patient studies to investigate the benefit of 7 Tesla for the human body.