Skip to main content
2025 SCMR Conference Main banner
Final Program


Schedule at a Glance
Back

Rapid Fire Abstracts

Rapid Fire 06- Machine Learning and Artificial Intelligence III

Generative Diffusion Motion Model for Improved Myocardial Strain Analysis of Cine Magnetic Resonance Images (RF_FR_327)

Friday, January 31, 2025
11:20 AM – 11:30 AM East Coast USA Time
Location: Blue Room PreFunction - Kiosk 1
Background: Myocardial strain analysis of routine cine balanced steady state free precession (bSSFP) cardiac magnetic resonance (CMR) images is important for evaluating cardiac function [1,2]. Despite strides made using deep learning to automate strain analysis, existing methods have limitations [3,4,5]. Recent research has harnessed strain-dedicated techniques, such as displacement encoding with stimulated echoes (DENSE) CMR, to train deep networks in predicting DENSE-comparable myocardial displacement from cine CMR [6,7]. However, these approaches rely heavily on myocardial contours rather than directly addressing the underlying deformation/motion patterns, which compromises the strain accuracy.

Methods: We utilize a newly developed generative model, termed latent motion diffusion model (LaMoD) from our research group [8], to predict highly accurate myocardial motion from standard CMR image videos, with the guidance of DENSE ground truth. Our model, LaMoD, includes two key components: (i) a registration network that extracts latent motion features from input DENSE contours, and (ii) a powerful probabilistic latent diffusion model [10,11] that learns the data distribution of latent myocardial motion features to reconstruct DENSE-quality displacements. Once trained, DENSE data is no longer required for testing. We evaluate our model's performance on standard cine CMR by evaluating segmental circumferential strain (Ecc) absolute error and analyzing the statistical agreement with DENSE ground-truth Ecc at end-systole. Our method is compared against state-of-the-art deep learning models of CMR strain analysis, including StrainNet [6], UNetR [12], and 3D TransUNet [13]. We also compare with a widely used commercial feature tracking software (SuiteHeart version 5.0.4; NeoSoft) for clinical reference. We include 741 DENSE CMR contour videos from 284 subjects (124 healthy volunteers, 160 patients with various heart diseases) for model training and 105 short-axis cine bSSFP CMR slices from 40 subjects for testing.

Results:

Fig. 2 highlights LaMoD's superior correlation with ground-truth DENSE data, achieving global and segmental correlation coefficients of R=0.91 and R=0.77, respectively, surpassing all baseline models (global R∈[0.83, 0.86], segmental R∈[0.57, 0.73]). This emphasizes LaMoD's accuracy and clinical robustness. Visual comparisons in Fig. 3 further demonstrate LaMoD's enhanced precision across both healthy volunteers and heart disease patients. Quantitatively, LaMoD achieves a mean strain error of 3.67 ± 1.16%, outperforming baselines: UNetR (4.41 ± 1.10%), TransUNet (4.29 ± 1.62%), StrainNet (4.41 ± 1.28%), and feature tracking (FT) (4.63 ± 1.50%).

Conclusion: Our developed model, LaMoD, provides a new way for accurate myocardial motion and strain prediction from standard CMR videos. It demonstrates superior performance compared to existing methods, potentially improving cardiac disease assessment and treatment planning without requiring additional DENSE scans. Future work will include explicit modeling of myocardial rotation over time, which further improves strain analysis in standard cine CMR by accounting for movements related to twist and torsion.


Acknowledgement: This work was supported by NIH 1R21EB032597.