Distortion correction of single-shot EPI enabled by deep-learning.

Area of Science:

• Magnetic Resonance Imaging (MRI)
• Medical Image Processing
• Deep Learning

Background:

• Echo planar imaging (EPI) is susceptible to geometric distortions.
• Accurate distortion correction is crucial for quantitative analysis in MRI, especially for diffusion-weighted imaging (DWI).
• Existing correction methods like field-mapping and top-up have limitations.

Purpose of the Study:

• To develop and validate a deep learning-based method for correcting distortions in single-shot EPI (SS-EPI) images.
• To utilize point-spread-function encoded EPI (PSF-EPI) as a reference for distortion correction.
• To compare the proposed method against established distortion correction techniques.

Main Methods:

• A 2D U-net deep neural network was trained using PSF-EPI images as targets for SS-EPI distortion correction.
• Anatomical T2-weighted turbo spin-echo (T2W-TSE) images were incorporated to enhance correction quality.
• The method was trained on healthy volunteers and tested on both healthy subjects and patients across different MRI platforms.

Main Results:

• The proposed deep learning method demonstrated superior EPI distortion correction compared to field-mapping and top-up methods.
• Corrected images closely approximated the quality of distortion-free PSF-EPI.
• Inclusion of T2W-TSE images improved SS-EPI distortion correction without introducing artifacts.
• Preliminary experiments indicated good generalization across patients and MRI platforms.

Conclusions:

• The deep learning-based approach effectively corrects EPI distortions, particularly in DWI.
• The method shows feasibility and potential for widespread clinical application in MRI.
• PSF-EPI serves as a valuable reference for training deep learning models for distortion correction.