Article, see p 589

Noninvasive cardiovascular imaging is fundamental to diagnosis, surveillance, risk stratification, and management of patients with hypertrophic cardiomyopathy (HCM). Although transthoracic echocardiography more frequently identifies or confirms the presence of HCM, cardiovascular magnetic resonance (CMR) is increasingly recommended to exclude phenocopies, screen family members, and make decisions about implantable cardiac defibrillator implantation to prevent sudden cardiac death.

Alongside improved delineation of cardiac structure and left ventricular (LV) myocardial function to measure LV myocardial mass, wall thickness, and systolic-diastolic function, CMR offers the potential to characterize the LV myocardial tissue and thereby identify the presence and extent of underlying disease processes. To date, much of this identification relies on gadolinium-based contrast agent (GBCA) administration to appreciate late gadolinium enhancement (LGE), which can detect extracellular pathology including replacement fibrosis and infarction, necrosis, extracellular edema, microvascular obstruction, hemorrhage, and infiltration.

Similar to the established relationship between myocardial scar and arrhythmogenicity in ischemic heart disease, there is evidence that significant myocardial scar observed with LGE portends a poor prognosis in HCM. As a result, international HCM guidelines^1,2 now include recommendations to consider CMR to identify LGE within the hypertrophied LV myocardium to improve risk stratification. Observing >15% of the LV myocardium with LGE portends an increased risk of future sudden cardiac death.

Acquiring and assessing LGE images during CMR have some limitations, including requirement for GBCA contrast administration, which lengthens the CMR procedure for the patient. The presence of LGE in those with HCM provides important prognostic information yet may underestimate potentially harmful underlying myocardial pathophysiology. HCM is characterized by diffuse LV myocardial interstitial fibrosis alongside myocardial hypertrophy and myocellular disarray, with islands of denser focal scar present only in more advanced disease. LGE imaging relies on regional heterogeneity in myocardial signal intensity to identify focal fibrosis but cannot discriminate diffuse fibrosis. Thus, correlation of LGE with histological fibrosis is weaker than in ischemic cardiomyopathy, where focal infarcts are generally surrounded by relatively unaffected LV myocardium. Also, LGE quantification in those with HCM is highly dependent on the thresholding technique used, making it a less robust tool for risk stratification.³

T1 mapping is a CMR technique that provides quantitative pixel-wise measurements of LV myocardial tissue composition without the requirement for GBCA with the signal influenced by both myocyte and extracellular interstitial compartments. T1 values are elevated in sarcomeric HCM (including phenotype-negative gene carriers), and the use of T1 mapping can discriminate HCM from other hypertrophic phenotypes.⁴ Unfortunately, there has been lack of standardization of T1 mapping, sequences, and thresholds for health and disease commonly overlap. Therefore, the initial promise of T1 mapping has not translated into routine clinical use as an independent diagnostic or prognostic biomarker aside from limited extreme phenotypes (eg, cardiac amyloidosis, iron deposition, and Fabry’s disease). T1 map analysis involves manual postprocessing to calculate the average T1 in a specific region of interest. Although this gives a measure of the fibrosis severity within a myocardial area, it provides little information about the affected myocardium volume, and reproducibility remains limited.

Over the last 5 years, automated analysis techniques using artificial intelligence and machine learning have become more prevalent in cardiovascular research (Figure). As the era of tracing of regions of interest comes to a close, the resultant improvements include accuracy and precision of artificial intelligence–contoured quantification of myocardial structure and function.⁵ In this issue of Circulation, Zhang et al⁶ describe the development of an artificial intelligence–based deep learning technology, “Virtual Native Enhancement” (VNE), from CMR noncontrast T1 maps and cine images, to produce synthetic images that closely resemble conventional LGE scar images. They trained and validated a neural network on the large international multicenter HCMR cohort (Hypertrophic Cardiomyopathy Registry),⁷ demonstrating strong agreement with LGE and better image quality. Furthermore, this new method demonstrated increased sensitivity for detecting relatively mild fibrosis regions, a feature particularly valuable in the context of sarcomeric HCM because it may facilitate earlier diagnosis during family screening or in prephenotypic gene carriers. The automated analysis used in VNE is based on the full-width-at-half-maximum quantification method, and T1 measurements were quality-controlled using phantom calibration. Together these advances could further improve measurement standardization, thereby permitting direct comparison between scans acquired at different times or on different scanners—of particular importance for detecting serial imaging changes or combining datasets from different centers for clinical research.

Figure. Timeline of cardiovascular magnetic resonance (CMR) biomarkers. A timeline depicts the evolution of CMR biomarkers, with key milestones indicated including the new potential role of VNE.⁶ A case example of a 26-year-old female with asymmetrical HCM is shown in short-axis (left image) and 4-chamber views (right image) demonstrating cine images with increased wall thickness of 3.7 cm in the inferoseptal wall (green boxes), LGE images with extensive patchy left ventricular enhancement in areas of hypertrophy (blue boxes), and diffusely elevated T1 values on T1 mapping (yellow boxes). AHA indicates American Heart Association; AI, artificial intelligence; ESC, European Society of Cardiology; ECV, extracellular volume; GBCA, gadolinium-based contrast agent; HCM, hypertrophic cardiomyopathy; HCMR, Hypertrophic Cardiomyopathy Registry; LGE, late gadolinium enhancement; MRI, magnetic resonance imaging; NMR, nuclear magnetic resonance; and VNE, virtual native enhancement.

The authors suggest that VNE analysis may obviate the requirement for contrast administration and could be applied across a broad range of cardiac pathologies. The potential advantages of widespread adoption of a noncontrast approach to CMR using automated analysis are clear as GBCA administration has some drawbacks including the following: (1) GBCAs are not well-suited for some patient groups, including those with severe renal dysfunction or allergies to GBCA, or where there are concerns related to gadolinium accumulation in the brain stem; (2) GBCAs require intravenous cannulation and physician supervision in case of allergy; and (3) LGE images must be acquired 5 to 10 minutes after GBCA administration, which, when combined with image acquisition, prolongs scan duration. Incorporating VNE imaging as opposed to GBCA-based LGE methods could reduce procedural length and provide faster and more reproducible postprocessing via automated analysis because observer variability may strongly influence CMR-derived measures such as wall thickness in HCM.⁸

There are still areas for future study and points worth highlighting. First, the improved visual image quality with VNE compared with the standard LGE images is unsurprising given that the VNE deep learning generator inputs a colocalized precontrast cine image alongside the T1 mapping data. CMR cine images have higher signal to noise than LGE images and significantly higher (≈5-fold) temporal resolution. This means the cardiac motion blurring seen in the longer LGE acquisitions is not found on the VNE (cine-based) images. In addition, as the authors themselves acknowledge, there needs to be comparison of VNE with contemporary improved LGE techniques. The sequences used for LGE identification in HCM have now been replaced by motion-corrected averaged sequences in many centers. Hence, some observed LGE image artifacts would not be found using newer sequences. Similarly, applying VNE to nonhypertrophic phenotypes may be more challenging; further model training is likely required to detect small subendocardial scars in thinned myocardium.

Second, although replacing LGE imaging with VNE imaging may reduce scan duration, this potential reduction may be less pronounced than suggested by the authors. Individual sequence acquisition times for T1 (required for VNE) and LGE are broadly similar, and although LGE images are acquired 10 minutes after contrast administration, most centers use that time interval to acquire the short axis cine images. Currently almost all CMR protocols include whole heart coverage for myocardial tissue characterization using LGE, including 3 long-axis images and a 8- to 12-image short-axis stack. The authors’ current study acquired only 3 short-axis T1 maps; hence, should VNE replace LGE imaging routinely in clinical practice, the protocol would need to be expanded significantly to attain equivalent whole-heart imaging.

Last, it is important to remember that T1 mapping (and hence VNE) and LGE CMR are not imaging equivalent myocardial disease processes, and precise correlation between the 2 techniques should not be expected. Given that T1 mapping (and hence VNE) derives signal not only from the interstitium (enabling measurement diffuse fibrosis) but also from the cardiac myocytes, it is likely to be able to detect acute myocardial injury including intracellular edema and inflammation with greater sensitivity than LGE imaging. For many cardiomyopathies, including HCM, there is increasing recognition that inflammation plays a key role in the underlying pathophysiology,⁹ and that this can be acute, chronic, or relapsing-remitting and likely impacts prognosis. The holy grail in HCM management is to transition toward stratifying patients by myocardial biology. This will provide personalized targeted therapies based not on clinical scenario (heart failure, outflow tract obstruction, arrhythmia) but on the underlying pathophysiology (myocyte hypertrophy, inflammation, fibrosis). Although the authors focus on replacing LGE imaging with VNE, further postprocessing of VNE and LGE images together to identify and quantify differences in signals between the 2 sequences may offer additional information related to the myocytes and intracellular disease processes including edema. Indeed, with the advent of multiparametric mapping and magnetic resonance fingerprinting, VNE has the potential to become a generative imaging technique providing novel insights into underlying myocardial biology.

In summary, CMR using LGE imaging has been transformative for noninvasive myocardial tissue characterization, thereby helping diagnose and risk-stratify cardiomyopathies including HCM. Zhang et al have provided valuable evidence that applying deep learning automated analysis techniques to standard noncontrast CMR sequences generates synthetic LGE images that correlate strongly with conventional LGE images, demonstrating their clinical utility in HCM. The VNE technique has the potential to improve feasibility and reduce overall CMR procedural time, which should further increase demand of this currently underutilized modality. The first major test for VNE will be its performance as a prognostic marker within the HCMR study itself. We await the outcome with much anticipation.

C.H.M. is directly and indirectly supported by the National Institute for Health Research Biomedical Research Centres at University College London Hospital and Barts Health National Health Service Trusts, United Kingdom. J.H.J. is supported by a grant from the Health and Environmental Sciences Institute and American Heart Association grant 19SLOI34580004. W.G.H. is supported by National Institutes of Health grants R01 HL118740, R01 CA199167, and R21 CA226960-01.

Disclosures None.

https://www.ahajournals.org/journal/circ

The opinions expressed in this article are not necessarily those of the editors or of the American Heart Association.

For Sources of Funding and Disclosures, see page 602–603.

文章，见第 589 页

无创心血管成像是肥厚型心肌病 (HCM) 患者诊断、监测、风险分层和管理的基础。尽管经胸超声心动图更频繁地识别或确认 HCM 的存在，但越来越多的人建议使用心血管磁共振 (CMR) 来排除表型、筛查家庭成员并做出关于植入式心脏除颤器植入的决定，以防止心源性猝死。

除了改善对心脏结构和左心室 (LV) 心肌功能的描绘以测量 LV 心肌质量、壁厚和收缩舒张功能外，CMR 还提供了表征 LV 心肌组织的潜力，从而确定潜在疾病过程的存在和程度. 迄今为止，这种鉴定大部分依赖于基于钆的造影剂 (GBCA) 给药来评估晚期钆增强 (LGE)，它可以检测细胞外病理，包括替代纤维化和梗死、坏死、细胞外水肿、微血管阻塞、出血和浸润.

与缺血性心脏病中心肌瘢痕和致心律失常性之间已建立的关系类似，有证据表明，用 LGE 观察到的显着心肌瘢痕预示着 HCM 的预后不良。因此，国际 HCM 指南^1,2现在包括考虑 CMR 以识别肥厚的 LV 心肌内的 LGE 以改善风险分层的建议。用 LGE 观察 >15% 的 LV 心肌预示着未来心脏性猝死的风险增加。

在 CMR 期间获取和评估 LGE 图像有一些限制，包括需要 GBCA 造影剂管理，这会延长患者的 CMR 程序。HCM 患者中 LGE 的存在提供了重要的预后信息，但可能低估了潜在有害的潜在心肌病理生理学。HCM 的特征是弥漫性 LV 心肌间质纤维化伴随着心肌肥厚和肌细胞紊乱，只有在更晚期的疾病中才会出现密集的局灶性瘢痕岛。LGE 成像依赖心肌信号强度的区域异质性来识别局灶性纤维化，但不能区分弥漫性纤维化。因此，LGE 与组织学纤维化的相关性弱于缺血性心肌病，其中局灶性梗塞通常被相对未受影响的 LV 心肌包围。此外，HCM 患者的 LGE 量化高度依赖于所使用的阈值技术，使其成为风险分层的稳健工具。³

T1 映射是一种 CMR 技术，它提供 LV 心肌组织成分的定量像素级测量，无需 GBCA，信号受心肌细胞和细胞外间质室的影响。肌节 HCM（包括表型阴性基因携带者）的 T1 值升高，使用 T1 定位可以将 HCM 与其他肥大表型区分开来。⁴不幸的是，健康和疾病通常重叠的 T1 映射、序列和阈值缺乏标准化。因此，除了有限的极端表型（例如，心脏淀粉样变性、铁沉积和法布里病）之外，T1 定位的最初前景并未转化为常规临床应用，作为独立的诊断或预后生物标志物。T1 地图分析涉及手动后处理以计算特定感兴趣区域的平均 T1。尽管这可以衡量心肌区域内的纤维化严重程度，但它提供的有关受影响心肌体积的信息很少，并且可重复性仍然有限。

在过去的 5 年中，使用人工智能和机器学习的自动分析技术在心血管研究中变得越来越普遍（图）。随着追踪感兴趣区域的时代即将结束，由此产生的改进包括人工智能的准确性和精确度——心肌结构和功能的轮廓量化。⁵在本期《流通》中，张等人⁶描述基于人工智能的深度学习技术“虚拟原生增强”（VNE）的发展，从 CMR 非对比度 T1 地图和电影图像，以生成与传统 LGE 疤痕图像非常相似的合成图像。他们在大型国际多中心 HCMR 队列（肥厚性心肌病登记处）上训练并验证了神经网络，⁷表现出与 LGE 的强烈一致性和更好的图像质量。此外，这种新方法在检测相对较轻的纤维化区域方面表现出更高的灵敏度，这一特征在肌节 HCM 的背景下特别有价值，因为它可以促进家庭筛查或表型基因携带者的早期诊断。VNE 中使用的自动分析基于半最大全宽量化方法，并且 T1 测量使用体模校准进行质量控制。这些进步一起可以进一步提高测量标准化，从而允许在不同时间或不同扫描仪上获得的扫描之间进行直接比较——这对于检测连续成像变化或组合来自不同中心的数据进行临床研究特别重要。

数字。 心血管磁共振 (CMR) 生物标志物的时间表。时间线描绘了 CMR 生物标志物的演变，并指出了关键里程碑，包括 VNE 的新潜在作用。⁶一个具有不对称 HCM 的 26 岁女性的病例示例显示在短轴（左图）和 4 腔室视图（右图）中，展示了下间隔壁中壁厚增加 3.7 厘米的电影图像（绿色框），LGE 图像在肥厚区域（蓝色框）具有广泛的左心室片状增强，T1 映射上的 T1 值弥漫性升高（黄色框）。AHA 表示美国心脏协会；AI，人工智能；ESC，欧洲心脏病学会；ECV，细胞外体积；GBCA，钆造影剂；HCM，肥厚型心肌病；HCMR，肥厚性心肌病登记处；LGE，晚期钆增强；MRI、磁共振成像；核磁共振，核磁共振；和 VNE，虚拟原生增强。

作者建议 VNE 分析可能不需要使用造影剂，并且可以应用于广泛的心脏病变。使用自动分析广泛采用非对比方法进行 CMR 的潜在优势是显而易见的，因为 GBCA 管理有一些缺点，包括以下内容：(1) GBCA 不太适合某些患者群体，包括那些患有严重肾功能不全或对GBCA，或存在与脑干中钆积累有关的问题；(2) GBCA 需要静脉插管和医生监督以防过敏；(3) GBCA 给药后 5 到 10 分钟必须采集 LGE 图像，这与图像采集相结合，延长了扫描时间。⁸

还有一些值得未来研究的领域和值得强调的地方。首先，考虑到 VNE 深度学习生成器在 T1 映射数据旁边输入共定位的预对比电影图像，与标准 LGE 图像相比，VNE 改进的视觉图像质量并不令人惊讶。CMR 电影图像比 LGE 图像具有更高的信噪比和显着更高（≈5 倍）的时间分辨率。这意味着在 VNE（基于电影）图像上没有发现在较长 LGE 采集中看到的心脏运动模糊。此外，正如作者自己承认的那样，需要将 VNE 与当代改进的 LGE 技术进行比较。在 HCM 中用于 LGE 识别的序列现在已被许多中心的运动校正平均序列所取代。因此，使用较新的序列不会发现一些观察到的 LGE 图像伪影。同样，将 VNE 应用于非肥大表型可能更具挑战性；可能需要进一步的模型训练来检测变薄的心肌中的小心内膜下疤痕。

其次，虽然用 VNE 成像代替 LGE 成像可能会减少扫描持续时间，但这种潜在的减少可能没有作者建议的那么明显。T1（VNE 所需）和 LGE 的单个序列采集时间大致相似，虽然 LGE 图像是在注射造影剂后 10 分钟采集的，但大多数中心使用该时间间隔来采集短轴电影图像。目前，几乎所有 CMR 协议都包括使用 LGE 进行心肌组织表征的全心脏覆盖，包括 3 个长轴图像和 8 到 12 个图像短轴堆栈。作者目前的研究仅获得了 3 个短轴 T1 图；因此，如果 VNE 在临床实践中常规取代 LGE 成像，则需要显着扩展该协议以实现等效的全心脏成像。

最后，重要的是要记住 T1 映射（以及 VNE）和 LGE CMR 不是成像等效的心肌疾病过程，并且不应期望这两种技术之间存在精确的相关性。鉴于 T1 映射（以及因此 VNE）不仅从间质（能够测量弥漫性纤维化）而且从心肌细胞获得信号，它可能能够以比 LGE 更高的灵敏度检测急性心肌损伤，包括细胞内水肿和炎症成像。对于包括 HCM 在内的许多心肌病，人们越来越认识到炎症在潜在的病理生理学中起着关键作用，⁹并且这可能是急性的、慢性的或复发-缓解并且可能影响预后。HCM 管理的圣杯是过渡到通过心肌生物学对患者进行分层。这将提供个性化的靶向治疗，而不是基于临床情况（心力衰竭、流出道阻塞、心律失常），而是基于潜在的病理生理学（肌细胞肥大、炎症、纤维化）。尽管作者专注于用 VNE 替代 LGE 成像，但进一步对 VNE 和 LGE 图像进行后处理以识别和量化 2 个序列之间的信号差异，可能会提供与肌细胞和细胞内疾病过程（包括水肿）相关的额外信息。事实上，随着多参数映射和磁共振指纹识别的出现，

总之，使用 LGE 成像的 CMR 对无创心肌组织表征具有变革性意义，从而有助于诊断和风险分层包括 HCM 在内的心肌病。Zhang 等人提供了宝贵的证据，表明将深度学习自动分析技术应用于标准的非对比 CMR 序列生成与传统 LGE 图像密切相关的合成 LGE 图像，证明了它们在 HCM 中的临床效用。VNE 技术有可能提高可行性并减少整体 CMR 程序时间，这应该会进一步增加对这种目前未充分利用的模式的需求。VNE 的第一个主要测试将是其作为 HCMR 研究本身的预后标志物的性能。我们满怀期待地等待着结果。

CHM 由伦敦大学学院医院国家卫生研究所生物医学研究中心和英国 Barts Health 国家卫生服务信托基金直接和间接支持。JHJ 得到了健康与环境科学研究所和美国心脏协会赠款 19SLOI34580004 的资助。WGH 由美国国立卫生研究院资助 R01 HL118740、R01 CA199167 和 R21 CA226960-01。

披露无。

https://www.ahajournals.org/journal/circ

本文中表达的观点不一定是编辑或美国心脏协会的观点。

有关资金来源和披露信息，请参见第 602-603 页。