Mitral Regurgitation Quantification Using Dual-venc 4D flow MRI and Deep learning

项目总结/摘要 二尖瓣返流（MVR）是最常见的瓣膜疾病之一，影响美国5%以上的人口。 人口及时准确地评估MVR对这些患者至关重要，因为MVR随时间推移而变化 未经治疗的重度MVR显著增加心力衰竭和死亡的风险。目前，超声心动图 （回声）是MVR的主要成像方式，其中MVR流量的定量在以下方面发挥着重要作用 确定疾病的严重程度。然而，回波的固有弱点（2D采集、单向速度 由于复杂的MVR血流动力学特征为高- 速度（4-6 m/s），非均匀（偏心/多/非全收缩射流）射流，动态变化 二尖瓣口形态心脏MRI（CMR）可用于基于以下指标间接量化MVR流量： 然而，在不同部位测量的每搏输出量的差异会放大每次测量的误差 由于减影，不适用于分流和/或多个瓣膜病变的患者。此外，本发明还 CMR和超声心动图之间的不一致性一直被报道，这表明需要准确和 可靠的定量技术。 4D flow MRI提供了对4D（3D+时间）心内血流的独特访问，从而能够“直接”量化 MVR射流动力学不受传统回波和CMR方法的限制。但临床 由于两个原因，这种方法的翻译仍然具有挑战性。一个是高速编码灵敏度 常规单静脉4D流动MRI需要4-6 m/s的（venc）来捕获高峰MVR流动射流速度。 这限制了4D流动MRI的速度动态范围，从而导致流动可视化差，并且增加了 流量量化不确定性。另一个是后处理需要人工繁琐的检测， 心动周期内3D全心脏中的MVR流动射流、多个时间范围内的平面放置和射流轮廓 限制了测量的再现性。该提案旨在通过开发一种快速的双- venc 4D flow MRI技术，针对MVR流速采集进行了优化，其次是深度学习技术 用于检测和分割4D MVR流射流，以实现MVR流量化过程的全自动化。的 具体目的是：（1）使用体外脉动MVR流动射流模型优化CS双腔4D流动MRI， (2)根据同一天采集的回波和CMR，验证60名MVR患者的双腔4D血流MRI 以及（3）开发深度学习网络以完全自动化MVR流量量化管道。 本项目将为MVR的临床评价提供一种可重现且准确的定量方法。我们 框架在成像、深度学习和临床应用方面享有多项创新。的经验教训 这应该适用于其他瓣膜损害的量化，从而大大扩大了影响 这个工作。

Project Summary/Abstract Mitral valvular regurgitation (MVR) is one of the most common valvular diseases affecting over 5% of the U.S. population. Timely and accurate assessment of MVR is crucial for these patients since MVR worsens over time and untreated severe MVR significantly increases risk of heart failure and death. Currently, echocardiography (echo) is the mainstay imaging modality for MVR where quantitation of MVR flow plays an instrumental role in determining disease severity. However, inherent weaknesses of echo (2D acquisition, 1-directional velocity measurements) limit quantification precision due to complex MVR hemodynamics characterized as a high- velocity (4-6 m/s), heterogeneous (eccentric/multiple/non-holosystolic jets) flow jets with dynamically changing mitral orifice morphology. Cardiac MRI (CMR) can be used to indirectly quantify MVR flow volume based on differences in stroke volumes measured at different sites, however, errors in each measurement are amplified due to subtraction and is inapplicable in patients with shunt flows and/or multiple valvular lesions. Further, discordance between CMR and echo has consistently been reported suggesting a need for an accurate and reliable quantitative technique. 4D flow MRI provides unique access to 4D (3D+time) intra-cardiac blood flow enabling “direct” quantification of MVR jet flSow dynamics free from limitations in conventional echo- and CMR-based methods. However, clinical translation of this approach remains challenging for two reasons. One is that a high velocity encoding sensitivity (venc) of 4-6 m/s is required for conventional single-venc 4D flow MRI to capture high peak MVR flow jet velocity. This limits velocity dynamic range of 4D flow MRI and thus, resulting in poor flow visualization and increased flow quantification uncertainty. The other is that post-processing requires manual and cumbersome detection of MVR flow jet in a 3D whole heart over a cardiac cycle, plane placement and jet contouring over many timeframes limiting measurement reproducibility. This proposal seeks to address these limitations by developing a fast dual- venc 4D flow MRI technique optimized for MVR flow velocity acquisition and second, a deep learning technique for detection and segmentation of 4D MVR flow jet to fully automate MVR flow quantification process. The specific objectives are: (1) to optimize CS dual-venc 4D flow MRI using in-vitro pulsatile MVR flow jet models, (2) to validate the dual-venc 4D flow MRI in 60 MVR patients against echo and CMR acquired on the same-day and (3) to develop a deep learning network to fully automate MVR flow quantification pipeline. This project will generate a reproducible and accurate quantitative approach for clinical evaluation of MVR. Our framework enjoys multiple innovations in imaging, deep learning, and clinical application. Lessons learned from this should be applicable to quantification of other valvular regurgitant lesions, thus greatly expanding the impact of this work.