Label-informed cardiac magnetic resonance image synthesis through conditional generative adversarial networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Scarcity of annotated medical data hinders AI research.
• High-quality synthetic medical images can address data limitations.

Purpose of the Study:

• Develop a novel framework for generating high-fidelity, diverse Cardiac Magnetic Resonance (CMR) images.
• Evaluate the utility of synthetic CMR images for deep learning-based cardiac cavity segmentation.

Main Methods:

• A two-module framework using mask-conditional Generative Adversarial Networks (GANs).
• Module 1: Segmentation for multi-tissue labels on real CMR images.
• Module 2: Synthesis to translate segmentation masks into realistic CMR images.

Main Results:

• Synthetic data trained models achieved performance comparable to real data.
• Augmenting real data with synthetic data improved Dice score (max 4%) and Hausdorff Distance (max 40%).
• Investigated impact of label quantity, training data size, and multi-vendor data.

Conclusions:

• The proposed framework successfully generates diverse and realistic synthetic CMR images.
• Synthetic data shows strong potential for data augmentation and replacement in medical image analysis.
• This approach can help overcome data scarcity challenges in cardiac imaging research.