Generative Models to Synthesize High-quality Displacement and Strain from Routine Cine Cardiac Magnetic Resonance Images

Background: Cardiac magnetic resonance (CMR) strain imaging adds clinical value in applications such as cardiotoxicity, cardiomyopathies, cardiac resynchronization therapy and others [1]. Displacement Encoding with Stimulated Echoes (DENSE) is widely considered a gold standard strain method [2-4]; however, it requires additional acquisitions, which is not ideal for an efficient clinical workflow. We have previously shown the feasibility of training artificial intelligence (AI) models (e.g., StrainNet [5]) using supervised learning from DENSE data where displacement and strain inference can be applied for the analysis of routine cine CMR [5, 6]. Nevertheless, these models require spatial and temporal pre-alignment of the cine and DENSE data - a process that can introduce artifacts and limit robustness and generalizability. This study extends and improves existing works by leveraging generative AI to learn the distribution of myocardial motion, enabling it to generalize across cine CMRs with wide variability in spatial and temporal resolution.

Methods: To improve the agreement between AI-predicted displacement and strain with reference DENSE, we propose a generative StrainNet model, GenStrain, to synthesize high-quality spatiotemporal displacement fields directly from cine CMRs. GenStrain includes two key components (Fig. 1): (1) a temporal latent registration network (TLRN) [7] that uses both DENSE and cine CMR data to extract spatiotemporal motion from segmented left ventricular (LV) contour videos; and (2) a conditional video diffusion model [8] to learn the distribution of DENSE displacements conditioned on TLRN-predicted motions combined with LV contours. The model was trained on 759 LV short-axis DENSE videos (15-30ms, ~2.41mm/pixel) from 157 subjects, with each video standardized to 48×48×20 (2D×t). Once trained, we validated the model on 40 cine videos (23-40ms, ~1.68mm/pixel) from cancer patients undergoing screening for cardiotoxicity (n=34) and healthy volunteers (n=6). Finally, we benchmarked GenStrain against existing methods [5, 6] for comparative evaluation.

Results: In Fig. 2, GenStrain achieves the lowest mean squared error (MSE) and mean absolute error (MAE) compared to DENSE-derived strain values, with lower whole-slice (0.08% MSE, 2.12% MAE) and segmental (0.14% MSE, 2.84% MAE) errors. Fig. 3 shows predicted displacement fields (top) and corresponding strain curves (bottom) compared with DENSE. GenStrain achieves the best agreement with DENSE ground truth, underscoring its ability to more accurately capture displacement and strain dynamics across the cardiac cycle.

Conclusion: We present GenStrain, a generative diffusion model to predict high-quality myocardial displacement and strain from routine cine MRI with improved agreement with DENSE. In contrast to previous approaches [5, 6]. GenStrain was trained on DENSE data with varied temporal and spatial resolutions, enabling robust generalization to cine MRI and demonstrating potential for broad clinical use.

Fig 1: Our proposed framework, GenStrain, consists of two main components: (1) an unsupervised temporal latent registration network which leverages both DENSE and cine CMR image data to extract spatiotemporal motion from segmented left ventricular (LV) contour videos; and (2) a conditional video diffusion model that learns the distribution of DENSE motion conditioned on the initial motion estimated by the registration network and the LV contours. During inference, GenStrain takes cine LV contour videos as input and generates displacement fields conditioned on registration-predicted heart motion.
Fig 2: The models were evaluated for whole-slice and segmental strain errors, measured by Percentage MSE and MAE. GenStrain achieved the lowest errors (whole-slice MSE/MAE: 0.08%/2.12%; segmental MSE/MAE: 0.14%/2.84%), followed by LaMoD (whole-slice: 0.10%/2.35%; segmental: 0.17%/3.21%) and StrainNet (whole-slice: 0.11%/2.53%; segmental: 0.27%/3.87%).
Fig 3: Representative examples of cine displacement fields and the corresponding whole-slice and segmental strain curves compared across all evaluated methods against the DENSE-derived ground truth.