An inline deep learning based free-breathing ECG-free cine for exercise cardiovascular magnetic resonance

Area of Science:

• Cardiovascular imaging
• Medical artificial intelligence
• Magnetic Resonance Imaging

Background:

• Exercise cardiovascular magnetic resonance (Ex-CMR) is a key stress test for coronary artery disease (CAD).
• Accelerated imaging techniques in Ex-CMR often cause significant aliasing artifacts.
• Current methods necessitate improvements for artifact reduction and real-time acquisition.

Purpose of the Study:

• To develop and evaluate a deep learning (DL)-based radial acceleration technique for free-breathing, electrocardiogram (ECG)-free real-time Ex-CMR.
• To assess the effectiveness of DL in suppressing artifacts in accelerated real-time cine images.
• To compare image quality and cardiac measurements between standard and DL-accelerated real-time Ex-CMR.

Main Methods:

• A 3D convolutional neural network was trained to suppress artifacts in simulated radial cine images.
• A prototype real-time radial sequence with an acceleration rate of 12 was used for inline DL reconstruction.
• Image quality was evaluated by three readers using a 4-point Likert scale, comparing standard cine and real-time radial cine images at rest and during stress.

Main Results:

• The DL model effectively reduced artifacts in real-time radial cine images, with 89.4% of rest images and 84.6% of stress images scoring moderate to minimal artifacts.
• Mean artifact scores for real-time images at rest (3.3 ± 0.7) and during stress (3.1 ± 0.6) indicated significant artifact suppression.
• While LV end-diastolic volume showed no significant difference, LV end-systolic volume and LV ejection fraction differed significantly between standard and real-time cine at rest.

Conclusions:

• This study demonstrates the feasibility of inline real-time cine using DL-based radial acceleration for Ex-CMR.
• The developed technique shows promise for improving image quality in accelerated Ex-CMR protocols.
• Further research can optimize DL algorithms for enhanced artifact reduction and clinical application in CAD diagnosis.