An end-to-end deep learning method for reconstructing SMS-PI accelerated musculoskeletal MRI.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Deep Learning (DL) accelerates musculoskeletal MRI by up to 6-fold, but further improvements in speed and generalization are needed.
• Novel approaches integrating Simultaneous Multislice (SMS) and Parallel Imaging (PI) are proposed to enhance DL-based reconstruction.

Purpose of the Study:

• To develop an advanced DL reconstruction framework combining SMS and PI for musculoskeletal MRI.
• To achieve accelerations of 8-fold and beyond while maintaining diagnostic image quality for clinical interpretation.

Main Methods:

• An End-to-End (E2E) DL framework was developed for reconstructing Turbo Spin Echo (TSE) MRI data acquired with SMS and PI acceleration.
• The framework integrates a novel DL network for joint slice regularization and embeds the SMS forward model into the DL architecture.
• The model was trained on over 200,000 MRI slices across various field strengths (1.5T-3T) and acquisition settings.

Main Results:

• The E2E DL model demonstrated superior performance compared to previous methods at 8-fold and 12-fold acceleration, based on PSNR and SSIM metrics.
• Radiological evaluation on prospectively acquired clinical scans confirmed comparable image quality and abnormality detection to standard, lower-acceleration acquisitions.

Conclusions:

• The study presents an E2E DL approach integrating slice separation for SMS acquisitions, advancing state-of-the-art reconstruction.
• Clinical-grade image quality was achieved at 8-fold acceleration, reducing MRI acquisition time by 27% in subjects.
• Preliminary findings indicate potential for further acceleration up to 12-fold, showcasing significant progress in DL-based MRI.