Synthetic Cardiac MRI Image Generation using Deep Generative Models
Organizations: Department of Computer Engineering, University of Peradeniya, Peradeniya, Sri Lanka · Department of Computer Engineering, University of Maryland, College Park, USA · Department of Holistic Systems, SimulaMet, Oslo, Norway
Abstract
Synthetic cardiac MRI (CMRI) generation has emerged as a promising strategy to overcome the scarcity of annotated medical imaging data. Recent advances in GANs, VAEs, diffusion probabilistic models, and flow-matching techniques aim to generate anatomically accurate images while addressing challenges such as limited labeled datasets, vendor variability, and risks of privacy leakage through model memorization. Maskconditioned generation improves structural fidelity by guiding synthesis with segmentation maps, while diffusion and flowmatching models offer strong boundary preservation and efficient deterministic transformations. Cross-domain generalization is further supported through vendor-style conditioning and preprocessing steps like intensity normalization. To ensure privacy, studies increasingly incorporate membership inference attacks, nearest-neighbor analyses, and differential privacy mechanisms. Utility evaluations commonly measure downstream segmentation performance, with evidence showing that anatomically constrained synthetic data can enhance accuracy and robustness across multi-vendor settings. This review aims to compare existing CMRI generation approaches through the lenses of fidelity, utility, and privacy, highlighting current limitations and the need for integrated, evaluation-driven frameworks for reliable clinical workflows.
Figures & tables
| Metric | Definition | Level (Pixel / High) | Application (Generation / Segmentation) | Frequency in Literature |
| Fréchet Inception Distance (FID) | Measures distribution similarity between real and generated image features; captures fidelity and diversity. | High (Feature / Distribution) | Image Generation (Standard metric for GANs / Diffusion models) | Very High |
| Structural Similarity Index Measure (SSIM) | Measures structural similarity and perceptual quality between two images. | Pixel-Level (Structural) | Image Generation; MRI Reconstruction | High |
| Multi-Scale SSIM (MS-SSIM) | Multi-scale variant of SSIM assessing structure at various resolutions. | Pixel / High (Multi-scale) | Image Generation (Structural fidelity) | High |
| Peak Signal-to-Noise Ratio (PSNR) | Measures pixel-level fidelity based on error magnitude. | Pixel-Level (Error) | MRI Reconstruction; Image Generation | High |
| Dice Similarity Coefficient (DSC) | Measures overlap between segmentation prediction and ground truth. | High (Spatial Overlap) | Segmentation Models; Utility of Synthetic Data | Very High |
| Hausdorff Distance (HD) | Measures boundary distance between two shapes/segmentations. | Pixel / High (Boundary Distance) | Segmentation Models (Boundary accuracy) | Medium |
| Dataset Name | Vendors / Sites | Number of Samples | Label Availability | Usefulness |
| M&Ms Challenge (Multi-Center, Multi-Vendor, Multi-Disease) | 6 hospitals across Spain, Canada, Germany; Vendors: Siemens, Philips, GE, Canon | 350 images [ 9 ] ; 150 annotated, 25 unannotated | Open-access; LV/RV/MYO labels at ED/ES annotated by clinicians | Most diverse CMR dataset; benchmark for generalization across vendors [ 13 ] ; used in image synthesis and multi-tissue segmentation studies [ 9 ] . |
| UK Biobank (UKB) | Single center; Siemens MAGNETOM Aera 1.5T | Over 5,000 subjects; some studies use 43,352 participants or 25,480 SAX cine sequences | Restricted access; LV/RV annotations at ED/ES | Used for large-scale cardiac function studies, phenotype-guided generation [ 40 ] , fairness correction, debiasing, and conditional CMR synthesis [ 41 ] . |
| ACDC Challenge | Single clinical center | 150 subjects | Public; LV/RV/MYO segmentation at ED/ES; includes NOR, DCM, HCM pathologies | Established segmentation benchmark [ 13 ] ; widely used for domain shift [ 33 ] and deep learning evaluation. |
| OCMR Dataset | Siemens Prisma 3T, Avanto 1.5T, Sola 1.5T | 53 fully sampled + 212 under-sampled (total 1383 k-space entries) | Open-access; full multi-coil k-space data (healthy subjects) | Used for DL reconstruction, acceleration studies [ 17 ] , and unconditional diffusion model training [ 8 ] . |
| M&Ms-2 Challenge | Multi-vendor; testing on Siemens, Philips, GE, Canon | 50 studies per vendor; extra 50 from unseen vendor (Canon) | Segmentation labels for LV/RV/MYO | Rigorous benchmark for generalizing to unseen moderate/severe cardiac pathologies [ 20 ] . |
| MS-CMRSeg Challenge | Multi-sequence challenge (LGE focus) | 45 subjects | Manual LV/RV/MYO labels by three observers; LGE-CMR images | Used for domain transfer, LGE-focused synthesis and segmentation methods [ 5 ] . |
| Reference | Generative Backbone | Dataset (Size / Limitation) | Conditioning (Mask / Covariate / Other) | Fidelity Evaluation | Utility Evaluation | Privacy Evaluation |
| UNCONDITIONAL GENERATION | ||||||
| Bubeck et al. (2022) [ 37 ] | UViT-B2 / DiT-B2 (Diffusion), MaskGIT (Autoregressive) | UK Biobank (11,360 subjects; large cohort ) | Unconditional generation | FID (38.78), KID | Benchmarking across spectrum; MaskGIT robust to masking | Not reported |
| Özdemir & Eroğul (2024) [ 8 ] | DDPM with asymmetric attention-enhanced U-Net | OCMR dataset ( limited data , restricted views) | Unconditional | FID (77.78), SSIM, MS-SSIM; radiologist accuracy 60% | Improves clinical downstream tasks; anatomical fidelity confirmed | Not reported |
| Tetralogy of Fallot Study (PG-GAN) [ 1 ] | Progressive GAN (PG-GAN), 256 256 output | Tetralogy of Fallot dataset (303 patients; limited ) | Unconditional latent-based generation | Expert review; Dice comparison against GT masks | Synthetic training achieves U-Net accuracy within 1% of real-data baseline | Synthetic images free of privacy constraints |
| CONDITIONAL GENERATION | ||||||
| Al Khalil / Amirrajab (2022–2023) [ 9 ] [ 3 ] | Conditional GAN with SPADE layers | M&Ms Challenge (multi-vendor/multi-disease; data scarcity ) | Multi-tissue segmentation maps (LV, RV, Myo, lung, fat, muscle) | SSIM, PSNR, NRMSE; SPADE superior to Pix2Pix/Pix2PixHD | Up to 4% Dice improvement; 40% HD reduction; replacement viability | Addresses privacy constraints in data sharing |