Variable-Density Single-Shot Fast Spin-Echo MRI with Deep Learning Reconstruction by Using Variational Networks

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Accelerated MRI acquisition is crucial for reducing scan times and improving patient comfort.
• Highly undersampled variable-density single-shot fast spin-echo (ssFSE) sequences offer speed but often suffer from lower image quality.
• Deep learning approaches show promise for enhancing reconstruction in accelerated imaging.

Purpose of the Study:

• To develop and evaluate a deep learning-based reconstruction method, variational network (VN), for highly undersampled ssFSE abdominal imaging.
• To compare the reconstruction speed and image quality of VN against conventional parallel imaging and compressed sensing (PICS).
• To assess the clinical feasibility of the VN approach for abdominal MRI.

Main Methods:

• A variational network (VN) was trained using data from 130 patients undergoing abdominal MRI with a 3.0-T imager.
• Coronal variable-density ssFSE sequences were acquired with 3.25x acceleration.
• Reconstruction performance was evaluated in 27 patients, comparing VN to PICS, with blinded radiologist assessment of image quality metrics.

Main Results:

• Variational network (VN) reconstruction was significantly faster (0.19s ± 0.04s) than PICS (5.60s ± 1.30s).
• VN demonstrated improved perceived signal-to-noise ratio (P = .01) and sharpness (P < .001).
• Overall image quality was superior with VN compared to PICS (P = .02), with no significant difference in contrast or artifacts.

Conclusions:

• Variational network (VN) reconstruction accelerates the process for accelerated ssFSE abdominal MRI.
• The VN approach enhances overall image quality, signal-to-noise ratio, and sharpness.
• VN represents a feasible and effective deep learning method for improving abdominal MRI reconstruction.