Introduction

CMR myocardial strain imaging is used diagnostically and prognostically for many types of heart disease. Feature tracking (FT) is a widely used and convenient method for strain MRI, as it applies post-processing algorithms directly to standard cine images to assess strain. It is, however, less accurate than strain-dedicated acquisitions like displacement encoding with stimulated echoes (DENSE)^1-4, especially for segmental strain. FT methods track myocardial contours rather than intramyocardial tissue because the myocardium presents uniform signal on cine MRI, lacking features to track. The intramyocardial motion is then (imperfectly) estimated using optical-flow based methods applied to the times series of endocardial and epicardial contours⁵. In contrast, DENSE directly measures intramyocardial tissue displacement; however, it requires additional acquisitions. As DENSE provides both myocardial contours and accurate intramyocardial tissue displacement information, we investigated the use of DENSE data to train deep networks to predict intramyocardial tissue motion from contour motion. This deep learning (DL) approach may provide the clinical convenience of FT and accuracy similar to DENSE.

Methods

Two approaches were developed and evaluated: (a) a two-step FlowNet2-based framework with a through-time correction network (TC-FlowNet2), and (b) a 3D Unet. TC-FlowNet2 framework: This network was built upon a successful optical-flow convolutional neural network (CNN) called FlowNet2⁶, which is widely used for frame-to-frame motion tracking of video imagery. We fine-tuned FlowNet2 using DENSE datasets, and we added a 3D through-time correction network to exploit the time dimension (Fig. 1A). 3D Unet framework: For this approach, a 3D Unet was trained to predict intramyocardial displacement from contour motion (Fig. 1B). For both approaches, during training, the inputs were a time series of myocardial contours derived from DENSE magnitude images and the ground truth data were DENSE tissue displacement measurements. Because DENSE and cine images at matched slice locations share similar motion patterns, we tested our trained model using contours derived from standard cine images (Fig. 1C). Data pre-processing for network training: We segmented the left-ventricular myocardium on DENSE and cine images, binarized the images by filling the myocardial area with 1 and the outside area and blood pool with 0, and cropped the images to a fixed size: N_x*N_y. Data augmentation was performed using 90° rotations. Cine images were scaled to match the resolution range of DENSE images. The input size for the FlowNet2-based network was two frames of endocardial and epicardial contours and the output of the DENSE-trained FlowNet2 was the frame-to-frame displacement field. The input of the through-time correction network was a stack of sequential displacements fields from DENSE-trained FlowNet2 with size of 2*N_x*N_y*N_t, where the factor of 2 accounts for displacements in two directions and N_t represents the number of temporal frames. The output was also size of 2*N_x*N_y*N_t. For the 3D Unet, the input size was N_x*N_y*N_t and the output size was 2*N_x*N_y*N_t. Datasets: Training datasets are described in Fig. 1D, and included a total of 60 volunteers and 42 patients with various pathologies such as left bundle branch block (LBBB), hypertrophic cardiomyopathy, dilated cardiomyopathy, coronary artery disease and hypertension. The model was tested on cine images of 10 volunteers and 18 patients using 3 short-axis views (base, mid-level and apex). For TC-FlowNet2, datasets were divided into two parts to separately train DENSE-trained FlowNet2 and the correction network, thus the testing dataset number (15 subject, 48 slices) was half the size as that used for the 3D Unet. Commercial feature-tracking (suiteHEART, Neosoft, WI) was also used to measure strain from cine images.

Results

Fig. 2 shows examples comparing TC-FlowNet2, 3D Unet and DENSE for computing end-systolic displacement and circumferential strain (E_cc) for a healthy subject and a LBBB patient. In these examples, both methods detect cardiac contraction in the healthy volunteer and stretching of the septum in the LBBB patient, but TC-FlowNet2 shows less contraction. Fig. 3 shows examples comparing commercial FT, TC-FlowNet2, 3D Unet and DENSE for computing global and segmental circumferential strain-time curves, with the 3D Unet showing better agreement with the ground truth (DENSE). Correlation plots and Bland-Altman plots (Fig. 4A, B) show that 3D Unet outperformed both TC-FlowNet2 and commercial FT for global and segmental E_cc. Also, as shown in Table 1, the intraclass correlation coefficient (ICC), coefficient of variation (CoV), and Pearson correlation coefficient (Pearson CC) showed that the 3D Unet provides the best agreement with DENSE, where the 3D Unet achieved ICC = 0.89 for global E_cc and ICC = 0.75 for segmental E_cc. Although TC-FlowNet2 showed good linearity relationship with DENSE, it has a relatively big bias, leading to its high Pearson CC but relatively low ICC.

Discussion and Conclusion

A 3D Unet, trained using DENSE datasets to predict intramyocardial motion from contour motion, outperformed both TC-FlowNet2 and commercial FT for the measurement of both global and segmental E_cc, for which DENSE data at matched locations served the reference standard.

Acknowledgements

NIH R01HL147104, UVA Ivy Biomedical Innovation Fund and AHA 2020AHAPRE0000203801.