GenStrain: Multi-Conditional Generative Diffusion Model To Synthesize High-quality Motion and Strain Maps From Cardiac Magnetic Resonance Images
Organizations: Department of Computer Science, University of Virginia, USA · Department of Biomedical Engineering, University of Virginia, USA · Department of Medicine, Division of Cardiology, University of Virginia, USA
Abstract
This paper presents GenStrain, a novel method for synthesizing high-quality cardiac motion-based strain maps from routinely collected cine cardiac magnetic resonance (CMR) images. Existing approaches typically rely on image registration to estimate motion (a sequence of displacements) from cine CMR images, but cine CMR presents very sparse trackable intramyocardial features, presenting challenges to estimate intramyocardial motion. An advanced imaging technique, Displacement Encoding with Stimulated Echoes (DENSE) CMR, is considered as the gold standard for strain assessment, but it requires additional acquisitions, which are not ideal for efficient clinical workflows. To address this gap, we introduce GenStrain, a novel framework designed to learn the data distribution of highly accurate cardiac motion captured by DENSE. In particular, we develop a diffusion process in motion space, where spatiotemporal continuity is carefully considered. To fully leverage both cine CMR images and their associated motion patterns, we introduce a multi-conditional conditioning strategy that integrates anatomical features and motion representations learned by a spatiotemporal registration model to guide the generation process. Experimental results demonstrate that our proposed model, GenStrain, outperforms existing strain analysis methods based on displacement estimates from both traditional registration approaches and state-of-the-art (SOTA) deep learning models. Our model represents a promising step toward more accurate cardiac function assessment and holds great potential for clinical translation. Our code is publicly available at https://github.com/Rishov-MIA/GenStrain.
Figures & tables
| Whole-slice MAE (%) | Segmental MAE (%) | ||||
| Modality | Method | All-frame | ES | All-frame | ES |
| DENSE | DeepStrain | – | – | – | – |
| TLRN | 5.72 3.41 ∗ | 6.55 4.10 ∗ | 6.56 3.28 ∗ | 8.06 3.83 ∗ | |
| TransUNet | 3.81 2.09 ∗ | 5.51 3.78 ∗ | 4.56 1.76 ∗ | 6.88 3.48 ∗ | |
| StrainNet | 2.92 1.67 ∗ | 4.56 2.84 ∗ | 3.68 1.36 ∗ | 5.51 2.35 ∗ | |
| Flow-matching | 2.73 2.25 ∗ | 3.69 3.34 ∗ | 3.69 2.04 ∗ | 4.95 3.00 ∗ | |
| Cohort | Method | ICC | Pearson CC | CV (%) | Bias 95% LOA |
| DENSE Whole-slice | DeepStrain | – | – | – | – |
| TLRN | 0.39 [0.28, 0.57] | 0.52 [0.41, 0.82] | 108.26 [48.53, 163.74] | 5.63 15.47 | |
| TransUNet | 0.65 [0.58, 0.72] | 0.79 [0.74, 0.84] | 44.73 [39.33, 51.07] | 3.32 7.40 | |
| StrainNet | 0.77 [0.72, 0.82] | 0.89 [0.85, 0.92] | 33.84 [28.99, 39.08] | 2.39 5.91 | |
| Flow-matching | 0.80 [0.70, 0.86] | 0.81 [0.72, 0.87] | 39.96 [31.90, 50.75] | 1.07 7.49 | |
| ControlNet | 0.80 [0.73, 0.85] | 0.85 [0.80, 0.89] | 35.81 [30.23, 41.97] | 1.87 6.43 |
| Condition | Whole-slice MAE (%) | Segmental MAE (%) | |||||
| Modality | Variant | Contour | Motion | All-frame | ES | All-frame | ES |
| DENSE | Contour only | 2.18 1.59 | 2.91 2.59 | 2.77 1.48 | 3.68 2.44 | ||
| Motion only | 2.25 1.56 | 3.05 2.70 | 2.92 1.40 | 3.97 2.54 | |||
| GenStrain | 2.14 1.56 | 2.81 2.43 | 2.73 1.46 | 3.68 2.26 | |||
| Cine | Contour only | 2.84 1.54 | 3.17 2.39 | 3.52 1.40 | 4.94 2.74 | ||
| Motion only | 3.09 1.63 | 3.47 2.59 | 3.79 1.50 | 5.17 2.90 | |||
| Whole-slice MAE (%) | Segmental MAE (%) | |||
| Method | All-frame | ES | All-frame | ES |
| Feature Tracking | 2.84 1.46 | 2.86 2.54 | 4.69 1.34 ∗ | 5.97 2.59 ∗ |
| GenStrain | 2.78 1.49 | 2.62 1.95 | 3.46 1.39 | 4.60 2.84 |
| Holm | 1.000 | 1.000 | ||
| Method | Feature Tracking | DeepStrain | TLRN | TransUNet | StrainNet | Flow- matching | ControlNet | LaMoD | GenStrain |
| Inference time (s) | 20.00 | 1.39 | 0.12 | 0.03 | 0.004 | 7.45 | 3.64 | 0.30 | 3.46 |
| Allocated GPU memory (GB) | – | 0.77 | 0.02 | 0.51 | 0.02 | 1.77 | 0.85 | 0.48 | 0.67 |
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
| Paired difference (GenStrain baseline) | |||||
| Metric | Baseline | Baseline (mean) | HL estimate [95% CI] | Effect size (win rate) | Holm |
| All-frame nEPE GenStrain: 0.478 | TLRN | 0.767 | -0.287 [-0.332, -0.244] | 0.99 (96%) | ∗ |
| TransUNet | 0.630 | -0.152 [-0.175, -0.127] | 0.97 (96%) | ∗ | |
| StrainNet | 0.574 | -0.104 [-0.126, -0.080] | 0.84 (88%) | ∗ | |
| Flow-matching | 0.570 | -0.088 [-0.118, -0.061] | 0.81 (86%) | ∗ | |
| ControlNet | 0.577 | -0.086 [-0.108, -0.065] | 0.91 (88%) | ∗ | |
| Paired difference (GenStrain baseline) | |||||
| Metric | Baseline | Baseline (mean) | HL estimate [95% CI] | Effect size (win rate) | Holm |
| Whole-slice MAE (%) GenStrain: 2.39 | TLRN | 5.95 | -3.38 [-4.14, -2.50] | 0.92 (88%) | ∗ |
| TransUNet | 3.85 | -1.65 [-2.15, -1.00] | 0.70 (82%) | ∗ | |
| StrainNet | 3.01 | -0.69 [-1.08, -0.27] | 0.48 (68%) | 0.023 ∗ | |
| Flow-matching | 2.96 | -0.52 [-0.80, -0.24] | 0.55 (74%) | ∗ | |
| ControlNet | 2.89 | -0.43 [-0.77, -0.16] | 0.47 (66%) | 0.023 ∗ | |
| Paired difference (GenStrain baseline) | |||||
| Metric | Baseline | Baseline (mean) | HL estimate [95% CI] | Effect size (win rate) | Holm |
| Whole-slice MAE (%) GenStrain: 2.91 | DeepStrain | 4.60 | -1.67 [-2.24, -1.08] | 0.82 (80%) | ∗ |
| TLRN | 5.21 | -2.35 [-2.80, -1.93] | 0.98 (95%) | ∗ | |
| TransUNet | 5.28 | -2.48 [-2.95, -1.90] | 0.95 (90%) | ∗ | |
| StrainNet | 3.74 | -0.84 [-1.18, -0.52] | 0.74 (82%) | ∗ | |
| Flow-matching | 3.92 | -0.97 [-1.47, -0.50] | 0.65 (75%) | ∗ | |
Medical Image Synthesis
Medical image synthesis creates, translates, reconstructs, and edits clinical images to fill missing data and support analysis.
See all 334 papers