Subject-Specific Convolutional Neural Networks for Accelerated Magnetic Resonance Imaging

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Magnetic Resonance Imaging Physics

Background:

• Magnetic Resonance Imaging (MRI) offers superior soft-tissue contrast without ionizing radiation, but suffers from long acquisition times.
• Parallel imaging techniques accelerate MRI scans by reconstructing images from undersampled k-space data using receiver coil arrays.
• Current parallel imaging methods often rely on linear shift-invariant convolutional kernels trained on limited autocalibration signal (ACS) data.

Purpose of the Study:

• To introduce Robust Artificial-neural-networks for k-space Interpolation (RAKI), a novel method for accelerated MRI using scan-specific convolutional neural networks (CNNs).
• To evaluate RAKI's performance in ultra-high resolution brain MRI and quantitative cardiac MRI across various acceleration factors.
• To assess RAKI's noise resilience and image reconstruction capabilities compared to existing parallel imaging techniques.

Main Methods:

• RAKI employs three-layer CNNs trained exclusively on scan-specific ACS data, eliminating the need for large external training datasets.
• The method performs k-space interpolation to reconstruct images from undersampled data.
• RAKI was validated on ultra-high resolution brain MRI and quantitative cardiac MRI datasets acquired at different acceleration rates.

Main Results:

• RAKI demonstrated improved noise resilience, particularly at high acceleration rates and low signal-to-noise ratio (SNR) conditions, outperforming conventional methods.
• Successful image reconstruction was achieved for quantitative cardiac MRI, even when a single CNN was applied to images with varying contrasts.
• The method showed robustness against overfitting to specific image content, maintaining performance across diverse imaging scenarios.

Conclusions:

• RAKI offers a promising approach for accelerating MRI scans by improving k-space interpolation using CNNs.
• The technique enhances noise performance and maintains image quality, making it suitable for a wide range of MRI applications.
• RAKI's ability to generalize across different contrasts and its improved noise resilience suggest significant potential for clinical adoption.