LAUSR

Cardiac magnetic resonance imaging (MRI) plays an important role in the differential diagnosis, risk stratification, and evaluation of the treatment effects of cardiovascular diseases. Myocardial strain analysis has been developed as a contrast-free quantitative method for the evaluation of myocardial deformation in different orientations, including longitudinal, circumferential, and radial strains. Among these, left ventricular global longitudinal strain (GLS) is considered the most sensitive parameter to detect early disease because subendocardial fibers are longitudinally directed and sensitive to ischemia. The left ventricular global circumferential strain (GCS) is also widely used in myocardial strain analysis and can reflect the extent of midwall or transmural damage of the myocardium. Impaired GCS is known to be an independent predictor of future cardiac events in patients with myocardial infarction, even after adjusting for left ventricular ejection fraction and the presence of late gadolinium enhancement. To date, several MRI techniques have been available for evaluating myocardial strain, including tagging MRI, displacement encoding with stimulated echoes (DENSE), strainencoded MRI, and feature tracking (FT) analysis using cine MRI images. As opposed to other MRI strain techniques, which require dedicated sequences, cine-based FT can provide quantitative assessment of myocardial strain in various settings and can be performed retrospectively using routine cine MRI data. However, a time-consuming analytical process, such as transferring the data to a dedicated workstation and manually drawing the region of interest around the left ventricle in each image, may limit the use of FT strain analysis in routine clinical practice. In this context, automated deep learning methods have been introduced into strain analysis using tagging, DENSE, and cine MRI images. These fully automatic methods have the potential to enhance the accuracy of epicardial and endocardial border tracing, data reproducibility, and analytical speed. In a single-center retrospective study presented in this issue of the Journal of Magnetic Resonance Imaging, the authors compared GCS and global radial strain (GRS) values between deep learning-based strain analysis and commercially available manual-based FT analysis in 47 healthy subjects and 533 cardiac patients. This study demonstrated that deep learningbased strain analysis showed excellent agreement with manual FT analysis (intraclass correlation coefficients: 0.82–0.90). This excellent agreement was also observed across two different MRI vendors: Philips Medical Systems and Siemens Healthcare. The present study is clinically relevant; however, caution should be exercised when interpreting the findings. First, the cine MRI data were retrospectively obtained from a single institution. The results of this study should be verified in a multicenter setting. Second, although the authors referred to GCS and GRS in the left ventricular short axis slices, they missed the evaluation of GLS, probably due to the lack of analysis of two-chamber and four-chamber views. Third, as there was no “gold standard” to assess the accuracy of strain measurements, a dedicated cardiac phantom study is needed to quantify and enhance the numerical accuracy of deeplearning-based strain analysis. In conclusion, this study represents an important step in applying cardiac MRI strain analysis to routine clinical practice. Various strain methodologies are now available; therefore, further well-designed prospective studies are needed to determine the difference in the total processing time and cost between deep learning-based and conventional semiautomatic myocardial strain analyses.