[論文レビュー] Synthetic Cardiac MRI Image Generation using Deep Generative Models

Ishan Kumarasinghe, Dasuni Kawya+4 more|arXiv (Cornell University)|Mar 25, 2026

Advanced Neural Network Applications被引用数 0

ひとこと要約

A comprehensive review comparing conditional and unconditional CMRI generation methods (GANs, VAEs, diffusion, flow-matching) across fidelity, downstream utility, and privacy, with emphasis on preprocessing, multi-vendor generalization, and evaluation frameworks.

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.

研究の動機と目的

• Motivation to overcome limited annotated CMRI data and privacy concerns in cardiac imaging.
• Evaluate how conditioning on segmentation masks improves anatomical fidelity.
• Assess the utility of synthetic CMRI for downstream segmentation and clinical tasks.
• Discuss privacy risks in generative CMRI and propose evaluation approaches.

提案手法

• Survey of generative model families (GANs, VAEs, diffusion, flow matching) for CMRI synthesis.
• Discussion of conditioning strategies, including mask-guided and vendor-specific conditioning.
• Analysis of preprocessing steps (rescaling, intensity normalization, ROI cropping, slice/temporal frame selection).
• Evaluation framework discussion focusing on fidelity metrics (FID, SSIM, PSNR, DSC, etc.) and downstream utility.
• Privacy audit considerations including membership inference attacks, frequency-calibrated attacks, and differential privacy implications.

実験結果

リサーチクエスチョン

• RQ1 What are the relative strengths and limitations of unconditional versus conditional CMRI generation methods for fidelity and clinical utility?
• RQ2 How do preprocessing, multi-vendor conditioning, and architectural choices affect cross-domain generalization in CMRI synthesis?
• RQ3 What are appropriate fidelity and utility metrics for evaluating CMRI generative models, and how should privacy be assessed and safeguarded?
• RQ4 Can segmentation-guided diffusion or flow-matching approaches achieve clinically meaningful anatomical fidelity for downstream tasks?
• RQ5 What privacy-preserving strategies are practical in CMRI generation without sacrificing image realism?

主な発見

• Diffusion models generally achieve superior fidelity in unconditional CMRI synthesis compared to GANs and VAEs.
• Conditional models using segmentation masks or multi-conditional prompts improve anatomical consistency and downstream task performance.
• Preprocessing steps like intensity normalization and ROI cropping are critical for cross-vendor robustness and synthesis quality.
• Evaluation should combine distribution-based metrics (e.g., FID) with application-driven tasks (e.g., segmentation accuracy) to avoid misinterpreting fidelity.
• Privacy evaluations via membership inference and nearest-neighbor analyses reveal potential leakage risks in synthesis, motivating post-hoc audits and regulatory-aligned safeguards.
• Various conditioning strategies (e.g., SPADE, ControlNet, diffusion with mask conditioning) enable finer control over anatomically plausible synthesis and cross-domain adaptation.