Synthetic Cardiac MRI Generation

01 - Overview

Research Abstract

Deep learning in cardiac MRI (CMR) is fundamentally constrained by both data scarcity and privacy regulations. This study systematically benchmarks three generative architectures: Denoising Diffusion Probabilistic Models (DDPM), Latent Diffusion Models (LDM), and Flow Matching (FM) for synthetic CMR generation. Utilizing a two-stage pipeline where anatomical masks condition image synthesis, we evaluate generated data across three critical axes: fidelity, utility, and privacy.

Our results show that diffusion-based models, particularly DDPM, provide the most effective balance between downstream segmentation utility, image fidelity, and privacy preservation under limited-data conditions, while FM demonstrates promising privacy characteristics with slightly lower task-level performance. These findings quantify the trade-offs between cross-domain generalization and patient confidentiality, establishing a framework for safe and effective synthetic data augmentation in medical imaging.

02 - Approach

System Pipeline

A two-phase pipeline: generative models produce synthetic CMR images from noise conditioned on anatomical masks, then outputs are rigorously evaluated across fidelity, utility, and privacy. Click each model tab to explore its architecture.

Phase 1 Generation

Real Mask

Gaussian Noise

→

Denoising Diffusion Probabilistic Model · Mask-Conditioned UNet

x_T
Noise

→

↓ DS
UNet

→

Mid
Block

→

↑ US
UNet

→

x̂₀
×T steps

Mask Conditioning

cross-attention injected at each diffusion timestep

Latent Diffusion Model · VAE Encoder–Decoder + Latent UNet

z_T
Noise

→

VAE
Encoder ↓

→

Latent
UNet

→

VAE
Decoder ↑

→

x̂₀
pixel space

Mask Conditioning

cross-attention in compressed latent space · reduced compute

Flow Matching · ODE Solver + Velocity UNet with ControlNet

x₀
Noise

→

ODE
Solver

→

Velocity
UNet

→

x₁
data space

ControlNet

Mask Embeddings

MultiScale Residuals

→

Synthetic CMR

↓

Phase 2 Evaluation

◑

Distribution

Based Evaluation

FIDKID

Statistical divergence between real and synthetic image distributions at the feature level

◎

Fidelity

Perceptual & Structural Quality

PSNRSSIM MS-SSIMLPIPS FIDKID

Full-reference & no-reference metrics for pixel, structural, and perceptual quality

◈

Utility

Downstream Segmentation Task

Real
Images+Masks

→

Seg
Model

←

Synthetic
+Masks

↓ evaluated on real test dataset

◻

Privacy

Patient Confidentiality

Nearest-Neighbour Analysis

L2SSIM LPIPSNNDR

Membership Inference Attack

FCREBinary Clf AUC

04 — Architectures

Generative Models Benchmarked

★ Best Overall

DDPM

Denoising Diffusion Probabilistic Model

Iterative denoising from Gaussian noise guided by learned score functions. Produces anatomically coherent, high-fidelity CMR images with strong segmentation utility under limited-data conditions.

FidelityHigh

UtilityHigh

PrivacyStrong

Efficient

LDM

Latent Diffusion Model

Diffusion process operating in a compressed latent space via a pre-trained encoder-decoder, enabling computationally efficient synthesis with competitive image quality at reduced resource cost.

FidelityGood

UtilityGood

PrivacyModerate

Novel Approach

Flow Matching

Continuous normalizing flow trained with simulation-free objectives. Demonstrates strong privacy characteristics with promising generalization, slightly below DDPM in task utility.

FidelityGood

UtilityModerate

PrivacyStrong

05 - Findings

Results & Comparative Analysis

Fidelity Evaluation

Evaluation Metric	Diffusion-DDPM	Diffusion-LDM	Flow Match
SSIM ↑	0.22	0.18	0.22
MS-SSIM ↑	0.36	0.33	0.40
PSNR ↑	10.67	9.95	11.44
FID ↓	72.52	95.17	108.32
KID ↓	0.04	0.08	0.098
LPIPS ↓	0.49	0.51	0.48

↓ lower is better · ↑ higher is better · Teal = best per metric

Evaluation of Segmentation Model

Training Setup	M&M Testing				ACDC Testing
Training Setup	Dice	IoU	HD95	ASD	Dice	IoU	HD95	ASD
M&M (Real)	0.90	0.84	2.99	1.04	0.91	0.83	2.89	0.94
ACDC (Real)	0.91	0.83	3.88	1.23	0.95	0.87	2.65	0.75
DDPM Full-Syn	0.87	0.80	5.77	1.79	0.87	0.80	6.28	1.78
DDPM ACDC-Syn	0.86	0.79	5.98	1.88	0.88	0.81	4.72	1.45
DDPM M&M-Syn	0.89	0.83	4.34	1.41	0.90	0.84	4.61	1.34
LDM Full-Syn	0.87	0.79	5.28	1.74	0.87	0.80	4.90	1.52
LDM ACDC-Syn	0.87	0.80	7.94	2.48	0.88	0.79	9.38	2.43
LDM M&M-Syn	0.89	0.82	4.17	1.40	0.88	0.82	2.26	1.53
FM Full-Syn	0.87	0.80	7.80	2.08	0.88	0.81	6.29	1.81
FM ACDC-Syn	0.82	0.73	8.75	2.90	0.85	0.76	7.74	2.19
FM M&M-Syn	0.85	0.82	5.04	1.64	0.89	0.82	5.73	1.67

Full-Syn: conditioned synthetic masks only · ACDC-Syn: conditioned on ACDC training masks · M&M-Syn: conditioned on M&M training masks · Teal = best among synthetic setups

Privacy Evaluation

Evaluation Metric	Diffusion-DDPM	Diffusion-LDM	Flow Match
Nearest Neighbor L2 ↑	12.0	10.5	19.0
LPIPS ↑	0.36	0.37	0.41
NNDR ↑	0.83	0.85	0.87
ROC_AUC (MIA) ↓	0.6029	0.580	0.6038

↑ higher is better for L2, LPIPS, NNDR · ↓ ROC-AUC closer to 0.5 indicates stronger privacy

FINDING 01 - Fidelity

DDPM achieves the best distribution-level fidelity with the lowest FID (72.52) and KID (0.04), while Flow Matching leads on pixel-level metrics PSNR (11.44) and MS-SSIM (0.40).

FINDING 02 - Utility

M&M-conditioned synthesis consistently outperforms other synthetic setups. DDPM M&M-Syn achieves the best ACDC Dice (0.90), while LDM M&M-Syn leads on M&M Dice (0.89) and HD95 (4.17).

FINDING 03 - Privacy

Flow Matching provides the strongest privacy with the highest L2 (19.0), LPIPS (0.41), and NNDR (0.87). LDM achieves the lowest MIA ROC-AUC (0.580), closest to the ideal 0.5.

FINDING 04 - Overall

All three models yield MIA AUC scores of 0.58-0.60, confirming robust resistance to re-identification attacks and validating the pipeline for safe synthetic data augmentation in medical imaging.