Accelerated coronary MRI with sRAKI: A database-free self-consistent neural network k-space reconstruction for...

Area of Science:

• Cardiovascular Imaging
• Magnetic Resonance Imaging
• Artificial Intelligence in Medical Imaging

Background:

• Accelerating Magnetic Resonance Imaging (MRI) acquisitions is crucial for improving patient comfort and reducing motion artifacts.
• Current parallel imaging techniques face limitations in handling arbitrary undersampling patterns and maintaining image quality at high acceleration rates.

Purpose of the Study:

• To develop and evaluate a novel deep learning-based reconstruction algorithm, sRAKI, for accelerating coronary MRI acquisitions.
• To enable arbitrary undersampling patterns in coronary MRI through coil self-consistency using subject-specific neural networks.

Main Methods:

• Introduced Self-consistent Robust Artificial-neural-networks for k-space interpolation (sRAKI), an iterative parallel imaging reconstruction method.
• Extended linear convolutions to nonlinear interpolation using convolutional neural networks (CNNs) trained on scan-specific autocalibrating signal (ACS) data.
• Enforced learned self-consistency and data-consistency to support random undersampling patterns in coronary MRI.

Main Results:

• sRAKI demonstrated superior performance over SPIRiT and l1-SPIRiT, reducing noise amplification and blurring artifacts, particularly at high acceleration rates (2-5x).
• Quantitative analysis showed significant improvements in normalized mean-squared-error and vessel sharpness compared to existing methods.
• Whole-heart coronary MRI data reconstructed with sRAKI exhibited the sharpest coronary arteries, with notable improvements in vessel sharpness.

Conclusions:

• sRAKI is a database-free, neural network-based reconstruction technique that accelerates coronary MRI with arbitrary undersampling.
• The method enhances noise resilience and image sharpness, outperforming traditional linear parallel imaging and l1 regularization techniques.