Real-time deep artifact suppression using recurrent U-Nets for low-latency cardiac MRI.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Interventions

Background:

• Real-time, low-latency Magnetic Resonance Imaging (MRI) is crucial for guiding cardiac interventions.
• Current iterative image reconstruction strategies for real-time MRI lead to prolonged reconstruction times.
• There is a need for faster reconstruction methods to improve the efficiency of cardiac interventions.

Purpose of the Study:

• To develop and evaluate a deep learning-based method for reconstructing highly undersampled radial real-time cardiac MRI data with low latency.
• To address the challenge of long reconstruction times in real-time MRI for cardiac interventions.

Main Methods:

• A 2D U-Net with convolutional long short-term memory layers was proposed to leverage spatial and temporal information.
• The network was trained on breath-hold CINE data and tested with retrospectively undersampled synthetic data across various acceleration rates and orientations.
• Prospective data were acquired and reconstructed in real-time during catheterization procedures in subjects.

Main Results:

• The proposed deep learning network demonstrated generalization to different acceleration rates and unseen orientations.
• Real-time reconstruction was achieved with low latency (39 ms), including deep artifact suppression (19 ms).
• In vivo imaging successfully depicted catheters, with image quality favorably comparing to existing reconstruction methods.

Conclusions:

• Deep learning-based artifact suppression was successfully implemented for non-Cartesian real-time interventional cardiac MRI.
• The developed method shows promise for time-critical applications in cardiac interventions, offering low latency and improved image quality.