Development and Application of StrainNet: Improved Myocardial Strain Analysis of Cine MRI by Deep Learning From DENSE
Wang, Yu, Biomedical Engineering - School of Engineering and Applied Science, University of Virginia
Epstein, Frederick, MD-BIOM Biomedical Eng PV-EVPP Office, University of Virginia
Myocardial strain imaging is used for the evaluation of multiple types of heart disease including the detection of chemotherapy-induced cardiotoxicity, for optimization of cardiac resynchronization therapy (CRT), for outcome prediction after myocardial infarction and for identification of subclinical cardiac dysfunction in obesity and diabetes. CMR methods such as myocardial tagging, displacement encoding with stimulated echoes (DENSE), and strain-encoded imaging acquire images specifically designed to measure intramyocardial deformation and strain, and can be referred to as strain-dedicated methods. Alternatively, feature tracking (FT) estimates strain from routine cine balanced steady-state free precession (bSSFP) images. Recent studies that evaluated both DENSE and FT in the settings of acute myocardial infarction and CRT found that DENSE outperformed FT for prognostication. Studies have shown that DENSE is reproducible for global and segmental strain, whereas FT has poor reproducibility for segmental strain. While DENSE provides well-validated and more predictive strain than FT, the time needed to acquire DENSE images may not always be compatible with an efficient clinical workflow. A strain method with performance similar to DENSE and the efficiency of FT would be ideal.
In this thesis, a novel deep learning workflow termed StrainNet was developed and validated to predict intramyocardial tissue motion and strain from myocardial contours. StrainNet was trained using 2D+t DENSE data and applied to cine bSSFP MR images, and validated on both healthy volunteers and patients from multiple sites. The performance of StrainNet was compared with commercial FT algorithm, with DENSE as the reference. To further improve the model performance in complex motion patterns especially motions with spatiotemporal long-term dependencies, we developed TransStrainNet, a transformer-based network combining the self-attention mechanisms for long-term dependencies and the locality properties of convolution, to capture both global and local patterns for improved intramyocardial motion estimation from contour motion. TransStrainNet was validated against StrainNet on general testing dataset and the subgroup of left bundle branch block (LBBB) with distinctive and complicated mechanical contraction patterns. In addition to technical development, TransStrainNet models were assessed on the prognostication of CRT patients and compared with commercial FT. The strain-based parameter (circumferential uniformity ratio estimated with singular value decomposition) from commercial FT and TransStrainNet would be calculated and used for 6-month response, 4-year survival predictions and risk stratifications.
In summary, StrainNet models provide accurate and convenient global and segmental strain analysis of routine cine MR images in healthy controls and patients with heart diseases, providing better prognostic performance for cardiac resynchronization therapy response and outcome predictions. These findings will facilitate greater use of strain cardiac MRI in research and in clinical settings.
PHD (Doctor of Philosophy)
Myocardial Strain Analysis, Deep Learning, DENSE, Motion Estimation, Feature Tracking
English
2024/04/18