1355787 - Multi-modality Strain Analysis: Generalization of CMR-trained StrainNet to Echocardiography

Myocardial strain imaging is used diagnostically and prognostically for many types of heart disease. Strain-dedicated CMR techniques such as displacement encoding with stimulated echoes (DENSE) and myocardial tagging serve as the gold standard for strain assessment and for validating other strain measurement techniques [1-3]. StrainNet is a recently-developed deep learning model trained with DENSE data that predicts intramyocardial motion from myocardial contours (Fig. 1A), and has been validated on cine MRI [4]. Since StrainNet can be applied to any image series that provides myocardial contour motion, we investigated the application of StrainNet to echocardiography images (StrainNet-echo). Leveraging a deep network trained from gold standard strain-dedicated DENSE CMR, we hypothesized that StrainNet would provide more accurate strain analysis of echocardiography than speckle tracking echocardiography (STE) and provide more interchangeable strain values between echo and CMR.

Methods:

Datasets: Eighteen patients with heart failure and left bundle branch block (LBBB) underwent CMR (with DENSE) and echocardiography examinations, and the imaging studies were performed within 2 days of each other to minimize time-dependent variations in myocardial function. Forty slices from these patients at 3 short-axis views (base, mid-level and apex) were analyzed. Data processing: Segmentation and speckle tracking of echocardiography images were performed using TomTec-Arena 2D Cardiac Performance Analysis 1.3.0.147 on the Agfa Healthcare Enterprise Imaging platform. For StrainNet analysis, echo images were scaled to match the mean spatial resolution of DENSE, and then were cropped to a fixed size. Images were binarized by filling the myocardial area with ones and the non-myocardial area with zeroes after LV segmentation using the TomTec STE LV contours (Fig. 1B).

Results:

Examples of end-systolic displacement and circumferential strain (E_cc) maps from StrainNet-echo and from CMR DENSE for a heart failure patient with LBBB are shown in Fig. 2A, and segmental strain-time curves corresponding to STE, StrainNet-echo and DENSE are shown in Fig. 2B. StrainNet-echo shows simultaneous stretching of the septal segments and contraction of the lateral wall, with good agreement with DENSE, whereas STE fails to show septal stretching, in disagreement with reference DENSE. While StrainNet-echo shows good agreement with STE for global E_cc, with an intraclass correlation coefficient (ICC) of 0.90, with regard to agreement with DENSE E_cc, correlation plots, Bland-Altman plots (Fig. 3A, B), ICC, Pearson correlation coefficient and coefficient of variation (Fig. 3C) all show that StrainNet-echo outperformed STE, especially for segmental strain.

Conclusion:

StrainNet predicts intramyocardial displacement and strain from echocardiography contours and shows good agreement with speckle tracking. StrainNet applied to echocardiography shows better agreement than STE with DENSE for both global and segmental E_cc.