Alternating Learning Approach for Variational Networks and Undersampling Pattern in Parallel MRI Applications | JoVE Visualize

Area of Science:

Medical Imaging
Computer Vision
Machine Learning

Background:

Accelerated parallel magnetic resonance imaging (MRI) aims to reduce scan times by acquiring fewer k-space samples.
Variational Networks (VN) are effective for reconstructing images from undersampled data, but their performance depends on the sampling pattern (SP).
Jointly optimizing SP and VN parameters is crucial for maximizing reconstruction quality in accelerated MRI.

Purpose of the Study:

To propose and evaluate an alternating learning approach for simultaneously learning sampling patterns (SP) and variational network (VN) parameters in accelerated parallel MRI.
To investigate variations of the alternating learning strategy, including different step types and learning rate schedules.
To assess the effectiveness of the learned SP-VN pairs for artifact removal and image quality improvement.

Main Methods:

An alternating learning framework was developed, iteratively optimizing SP and VN parameters using bias-accelerated subset selection and ADAM optimizer.
Four variations of the alternating approach were explored, incorporating monotone/non-monotone steps and learning rate reduction.
The learned SPs and VNs were evaluated on brain and knee datasets at acceleration factors (AF) from 2 to 20, comparing against other methods.

Main Results:

The proposed alternating learning approach demonstrated significant improvements in image quality, with gains of 1% to 62% over the next best method.
Visual inspection and quantitative metrics (RMSE, SSIM, HFEN) confirmed superior performance across various AFs and datasets.
The learned SPs captured fewer k-space samples, and the VNs effectively removed undersampling artifacts, showing stable performance across different initial conditions.

Conclusions:

The proposed alternating learning method effectively learns pairs of sampling patterns and variational networks for accelerated parallel MRI.
Improvements stem from both learned sampling density and precise sample positioning in k-space.
This approach enhances 3D Cartesian accelerated parallel MRI applications by enabling faster scans with high-quality image reconstruction.