Editorial for “Improved Quantification of Myocardium Scar in Late Gadolinium Enhancement Images: Deep Learning Based Image Fusion Approach”
Editorial for “Improved Quantification of Myocardium Scar in Late Gadolinium Enhancement Images: Deep Learning Based Image Fusion Approach”
复制标题
“晚期钆增强图像中心肌疤痕的量化改进：基于深度学习的图像融合方法”的社论
DOI：
10.1002/jmri.27619
复制
发表时间：
2021
影响因子：
4.4
通讯作者：
Suwa Kenichiro
中科院分区：
文献类型：
--
作者：
岡田健志;小関正博;西田誠;尾松卓;田中克尚;乾洋勉;冠野昂太郎;嵯峨礼美;大濱透;石原光昭;鯨岡健;服部浩明;増田大作;山下静也;坂田泰史;Suwa Kenichiro
肥厚型心肌病（HCM）的特征是左心室（LV）显著肥厚并伴有纤维化瘢痕。心血管磁共振（CMR）中的晚期钆增强（LGE）成像是一种潜在的诊断工具;此外，使用LGE评估瘢痕量化对于风险分层至关重要。迄今为止，已报道了LGE与室性心律失常、心脏骤停和死亡率的相关性（1）。然而，LGE的定量主要依赖于手动分割，因为心肌的多重结构，包括与肥大乳头肌和小梁的不清楚的内膜边界。此外，LGE和周围组织之间的对比是不明确的。因此，在LGE数据分割中，需要一种具有高再现性的劳动力较少的技术。为了克服上述问题，引入了计算机辅助以促进LGE数据的手动分割，其中强度通常被阈值化为距离远端心肌的平均强度的固定数量的标准偏差（2），或者选择用户选择的超增强区域内的最大强度的一半作为固定强度阈值（3）。这些方法已被广泛用于临床实践，并有助于疤痕分割;然而，LV本身的分割在很大程度上仍然依赖于手动分割。新的自动图像分割技术（4）已经开发出来，基于深度学习（DL）的图像处理已经成为最近研究的主流。通常，利用基于卷积神经网络的算法，并且将各种原始布置应用于卷积层、池化层和全连接层以获得更好的性能。在Moccia等人的一项研究中，(5)，与直接从LGE图像分割相比，在手动预定义的LV心肌中使用全卷积神经网络进行瘢痕分割实现了更高的性能，中值Dice相似系数（DSC）为71.25%，
Hypertrophic cardiomyopathy (HCM) is characterized by a significantly hypertrophic left ventricle (LV) with fibrotic scarring. Late gadolinium enhancement (LGE) imaging in cardiovascular magnetic resonance (CMR) is a potential diagnostic tool; furthermore, the assessment of scar quantification using LGE is crucial for risk stratification. To date, the associations of LGE with ventricular arrhythmia, cardiac arrest, and mortality have been reported (1). However, the quantification of LGE relies mainly on manual segmentation because of the multiplex structure of the myocardium, including an unclear endocardial boundary with hypertrophic papillary muscles and trabeculations. Furthermore, the contrast between LGE and surrounding tissues is ambiguous. Therefore, a technique that is less laborious with high reproducibility is warranted in LGE data segmentation. To overcome the aforementioned issues, computer assistance was introduced to facilitate the manual segmentation of LGE data, wherein the intensities are typically thresholded to a fixed number of standard deviations from the mean intensity of the remote myocardium (2) or half of the maximum intensity within a user-selected hyper-enhanced region is selected as the fixedintensity threshold (3). These methods have become widely used in clinical practice and have helped scar segmentation; however, the segmentation of the LV itself still largely relies on manual segmentation. Novel automatic image segmentation techniques (4) have been developed, and deep learning (DL)-based image processing has become the mainstream of recent studies. In general, a convolutional neural network-based algorithm is utilized, and various original arrangements to convolutional layers, pooling layers, and fully connected layers are applied for better performance. In a study by Moccia et al.(5), higher performance was achieved using a fully convolutional neural network in scar segmentation in the manually predefined LV myocardium compared to direct segmentation from LGE images with a median Dice similarity coefficient (DSC) of 71.25%,
影响因子：
24
作者：
Amado, LC;Gerber, BL;Lima, JAC
通讯作者：
Lima, JAC
影响因子：
19.7
作者：
Fahmy, Ahmed S.;Neisius, Ulf;Nezafat, Reza
通讯作者：
Nezafat, Reza