Optimized reconstruction of undersampled Dixon sequences using new memory-efficient unrolled deep neural networks:...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuromuscular Diseases

Background:

• Accurate fat fraction (FF) quantification in muscles via quantitative MRI is crucial for monitoring neuromuscular diseases.
• MRI acquisition undersampling is employed to shorten scan times, but can impact image quality.
• Developing efficient reconstruction methods is vital for clinical utility.

Purpose of the Study:

• Introduce novel unrolled neural networks for reconstructing undersampled 3D MRI data.
• Aim to maintain accurate FF quantification while reducing reconstruction time and memory footprint.
• Enhance the clinical applicability of quantitative MRI in neuromuscular disease assessment.

Main Methods:

• Utilized a simplified Half U-Net architecture combined with unrolled networks (VarNet, DIRCN).
• Trained and evaluated on 3D MRI Dixon thigh acquisitions from patients and controls.
• Applied retrospective undersampling with acceleration factors of 4 and 8.

Main Results:

• Novel networks maintained reconstruction, biomarker assessment, and segmentation quality.
• Achieved significant reductions in memory usage (16-24%) and reconstruction time (17-21%).
• At acceleration factor 8, HalfVarNet and HalfDIRCN demonstrated high SSIM and low MSE, with minimal relative FF error.

Conclusions:

• The proposed method offers time and memory-efficient reconstruction of undersampled 3D MRI.
• Demonstrates potential for improved clinical application of quantitative MRI in neuromuscular disease diagnosis and monitoring.