Improved reconstruction for highly accelerated propeller diffusion 1.5 T clinical MRI

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Diffusion MRI

Background:

• Propeller fast-spin-echo diffusion magnetic resonance imaging (FSE-dMRI) is crucial for diagnosing Cholesteatoma.
• Clinical 1.5 T MRI often yields low signal-to-noise ratio (SNR) in FSE-dMRI, necessitating longer scan times with increased signal averaging (number of excitations, NEX).

Purpose of the Study:

• To enhance SNR and accelerate PROPELLER FSE-dMRI on a 1.5 T clinical scanner using Locally Low Rank (LLR) reconstruction and Convolutional Neural Networks (CNNs).
• To reduce scan time while maintaining or improving diagnostic image quality for Cholesteatoma detection.

Main Methods:

• A Residual U-Net (RU-Net) architecture was developed and trained using 1-NEX FSE-dMRI data as input.
• The network was trained to predict 2-NEX images reconstructed with Locally Low Rank (LLR) constraints.
• Brain scans from healthy volunteers and patients with Cholesteatoma were utilized for training and validation.

Main Results:

• Offline LLR reconstruction improved image quality by suppressing noise and revealing small structures compared to online reconstruction.
• RU-Net further enhanced image quality, achieving a 18.87% increase in PSNR, 2.11% in SSIM, and a 53.84% reduction in NRMSE compared to LLR.
• RU-Net demonstrated a significant speed improvement, reconstructing images approximately 1500 times faster than LLR (0.03s vs. 47.59s per slice).

Conclusions:

• LLR reconstruction effectively boosts SNR in propeller FSE-dMRI.
• RU-Net significantly improves the performance of propeller FSE-dMRI in terms of PSNR, SSIM, and NRMSE.
• The RU-Net method enables a 2x reduction in scan time by utilizing 1-NEX data and offers a substantial speed advantage over LLR reconstruction.