Fast machine learning image reconstruction of radially undersampled k-space data for low-latency real-time MRI

Area of Science:

• Medical Imaging
• Machine Learning
• Image Reconstruction

Background:

• Fast data acquisition and reconstruction are crucial for real-time magnetic resonance (MR) imaging.
• Applications like MR-guided interventions require high temporal resolution.

Purpose of the Study:

• To develop and evaluate a machine learning (ML) model for accelerating the reconstruction of radially undersampled 2D k-space MR data.
• To assess the ML model's performance against conventional reconstruction methods in terms of image quality and speed.

Main Methods:

• A k-space-to-image ML model using a single fully connected linear layer for radial to Cartesian interpolation, followed by inverse fast Fourier transform.
• Training on synthetic natural images and evaluation on synthetic data (R=2-10) and real MR data from two systems (0.35T and 1.5T).
• Comparison with non-iterative zero-filling non-uniform fast Fourier transform (NUFFT) and compressed sensing (CS) reconstructions.

Main Results:

• The ML model achieved lower median mean squared error (MSE) than NUFFT on synthetic data.
• ML reconstruction times were faster than NUFFT and substantially faster than CS.
• The ML model demonstrated generalizability on real MR data, showing significantly shorter reconstruction times compared to conventional methods.

Conclusions:

• The proposed ML model effectively reconstructs MR data with reduced artifacts and extremely short reconstruction times.
• This ML approach is well-suited for rapid, low-latency, real-time MR applications.
• The model offers a promising alternative to conventional methods for accelerated MR image reconstruction.