Segmentation-Guided Diffusion for Free-Breathing Cardiac Magnetic Resonance Image Restoration.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Cardiac magnetic resonance imaging (CMR) is essential for cardiac function assessment but requires breath-holding.
• Free-breathing (FB) CMR is an alternative but suffers from motion artifacts, impacting diagnostic accuracy.
• Deep generative models show potential for artifact correction, but image fidelity remains a concern.

Purpose of the Study:

• To evaluate the impact of segmentation masks in diffusion models for enhancing anatomical structure preservation in image-conditioned generation.
• To improve the quality and diagnostic reliability of free-breathing cardiac magnetic resonance imaging.

Main Methods:

• Trained and evaluated a standard diffusion probabilistic model (DDPM) and a segmentation-guided DDPM on a public CMR dataset.
• Applied the models to restore motion-artifacted FB CMR data from a local hospital.
• Quantitatively assessed image quality using signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) in ventricular regions.

Main Results:

• The segmentation-guided DDPM significantly improved image quality compared to the standard DDPM.
• Higher SNR and CNR were achieved in both left and right ventricular regions using the guided approach.
• The segmentation-guided method demonstrated superior preservation of anatomical structures in reconstructed CMR images.

Conclusions:

• Segmentation-guided diffusion models effectively enhance anatomical structure preservation in image-conditioned generation for CMR.
• This approach offers a promising solution for improving the quality and diagnostic accuracy of free-breathing cardiac magnetic resonance imaging.
• The findings suggest a valuable method for overcoming motion artifacts in cardiac MRI, benefiting patients unable to hold their breath.