Balancing Fidelity, Utility, and Privacy in Synthetic Cardiac MRI Generation: A Comparative Study

Madhura Edirisooriya; Dasuni Kawya; Ishan Kumarasinghe; Isuri Devindi; Mary M. Maleckar; Roshan Ragel; Isuru Nawinne; Vajira Thambawita

cs.CVMar 4, 2026

Balancing Fidelity, Utility, and Privacy in Synthetic Cardiac MRI Generation: A Comparative Study

Authors: Madhura Edirisooriya, Dasuni Kawya, Ishan Kumarasinghe, Isuri Devindi, Mary M. Maleckar, Roshan Ragel, Isuru Nawinne, Vajira Thambawita

Organizations: University of Peradeniya, Sri Lanka · University of Maryland, College Park, USA · Tulane University School of Medicine, USA; Simula Research Laboratory, Norway · SimulaMet, Norway

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.

Figures & tables

Study	Model	U	F	P	CDU	LDU
[ 9 ]	Diffusion (SD-DM)	$\checkmark$	$\checkmark$	–	$\checkmark$	$\checkmark$
[ 10 ]	Diffusion (Attention)	–	$\checkmark$	–	–	$\checkmark$
[ 11 ]	Diffusion (DiffuSeg)	$\checkmark$	$\checkmark$	–	$\checkmark$	$\checkmark$
[ 12 ]	LDM vs GAN	$\checkmark$	$\checkmark$	–	$\checkmark$	$\checkmark$
[ 1 ]	Diffusion (SPADE)	$\checkmark$	$\checkmark$	–	$\checkmark$	$\checkmark$
[ 4 ]	Diffusion (MONAI)	–	$\checkmark$	–	–	–

TABLE I : Comparison of Generative Approaches Across Utility, Fidelity, and Privacy

Fig. 1 : Overview of the proposed synthetic cardiac MRI generation and evaluation framework. The pipeline consists of two main stages: (1) segmentation mask generation using a diffusion-based model, followed by (2) mask-conditioned image synthesis using DDPM, LDM, and Flow Matching. The generated images are evaluated across three dimensions: fidelity (PSNR, SSIM, MS-SSIM, LPIPS, FID, KID), downstream utility via cross-dataset cardiac MRI segmentation, and privacy through nearest-neighbor analysis and membership inference attacks.

Fig. 2 : Denoising Process. An image is shown each 200 timesteps

Feature	ACDC	M&Ms
Total Subjects	150	375
Centres	Single Centre	Six Centres
Vendors	Single (Siemens, protocol-specific)	Four (Siemens, Philips, GE, Canon)
Countries	1 (France)	3 (Spain, Canada, Germany)
Target Labels	LV, RV, MYO (ED & ES phases)	LV, RV, MYO (All the phrames)
Standard Utility	To test model performance on high-quality unseen data	Benchmarking cross-vendor generalisation

TABLE II : Comparison of ACDC and M&Ms Cardiac MRI Datasets

Fig. 3 : CMRI synthesis results. Top row: input segmentation masks. Subsequent rows: synthetic images generated by DDPM, LDM, and FM respectively

Structure	Feature	Real Mean	Gen Mean	p-value
LV	Pixel_Pct	5.1577	6.5380	0.0156
LV	Area	845.04	875.31	0.9084
LV	Roundness	0.9009	0.7160	$1.35\times 10^{-40}$
LV	Eccentricity	0.4929	0.4728	0.1112
LV	Solidity	0.9644	0.9001	$1.35\times 10^{-40}$
RV	Pixel_Pct	5.6487	5.9424	0.0061

TABLE III : Statistical Comparison Between Real and Generated Masks

Panel of Figure 7 — Fig. 4 : Comparison of geometric shape metrics between real and synthetic cardiac masks. (a) Area reflects the size of the segmented structure, (b) Eccentricity measures elongation, (c) Roundness evaluates circular compactness, and (d) Solidity assesses boundary smoothness and structural integrity. Overall, the real and synthetic distributions show strong alignment across most structures, with only minor deviations in (c) roundness and (d) solidity, indicating that the generated masks largely preserve anatomical geometry as assessed by (a)-(d).

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

TABLE IV : Results of Fidelity Evaluation

Training Setup	Testing
	M&M				ACDC
	Dice	IoU	HD95	ASD	Dice	IoU	HD95	ASD
M&M (Real)	0.90	0.84	2.99	1.04	0.91	0.85	2.89	0.94
ACDC (Real)	0.90	0.83	3.88	1.23	0.93	0.88	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	1.89	0.88	0.81	4.72	1.45

TABLE V : Evaluation of Segmentation Model

Evaluation Metric	Diffusion-DDPM	Diffusion-LDM	Flow Match
Nearest Neighbor (L2 Distance)	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

TABLE VI : Results of Privacy Evaluation. ROC-AUC values are from Membership Inference Attack.

Explore similar work

Synthetic Cardiac MRI Image Generation using Deep Generative Models
Mar 25, 2026Ishan Kumarasinghe, Dasuni Kawya, Madhura Edirisooriya +3Synthetic Data Generation Medical Image Synthesis
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
Metadata-Aware Adaptation of a Generative Foundation Model for Conditional CMR Synthesis
Aug 25, 2026Marc Rodríguez, Grzegorz Skorupko, Nay Aung +3Medical Image Synthesis Conditional Diffusion Models
Facultat de Matem`atiques i Inform`atica, Universitat de Barcelona, Spain · Barcelona Artificial Intelligence in Medicine Lab (BCN-AIM), Facultat de Matem`atiques i Inform`atica, Universitat de Barcelona, Spain · William Harvey Research Institute, NIHR Barts Biomedical Research Centre, Queen Mary University London, Charterhouse Square, London, UK +2
CardioDiT: Latent Diffusion Transformers for 4D Cardiac MRI Synthesis
Mar 26, 2026Marvin Seyfarth, Sarah Kaye Müller, Arman Ghanaat +7Medical Image Synthesis Latent Diffusion Models
Institute for Artificial Intelligence in Cardiovascular Medicine, Medical Faculty of Heidelberg University, Heidelberg University, Germany · Department of Cardiology, Angiology, Pneumology, Heidelberg University Hospital, Heidelberg, Germany · DZHK (German Centre for Cardiovascular Research), Partner Site Heidelberg/Mannheim, Heidelberg, Germany +4

Mar 25, 2026J/K move · Enter open · S save

Synthetic Cardiac MRI Image Generation using Deep Generative Models

Ishan Kumarasinghe, Dasuni Kawya, Madhura Edirisooriya, Isuri Devindi+2

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

Synthetic Data Generation Medical Image Synthesis Cardiac MRI Diffusion-Based Image Generation Synthetic Data Evaluation

Open paper AskPDF arXiv

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.

Abstract

Explore similar work

Synthetic Cardiac MRI Image Generation using Deep Generative Models

Metadata-Aware Adaptation of a Generative Foundation Model for Conditional CMR Synthesis

CardioDiT: Latent Diffusion Transformers for 4D Cardiac MRI Synthesis