Deep flow-net for EPI distortion estimation.

Area of Science:

• Medical Imaging
• Neuroimaging
• Machine Learning

Background:

• Geometric distortions in echo-planar imaging (EPI) are a significant challenge in functional and diffusion MRI.
• These distortions, caused by off-resonant spins, complicate accurate image analysis.
• Current methods like the blip up/down approach often rely on iterative techniques to estimate and correct these distortions.

Purpose of the Study:

• To explore the use of a deep convolutional network for estimating geometric distortion fields in EPI.
• To develop a faster and accurate method for correcting EPI distortions compared to traditional iterative approaches.
• To enable widespread application of distortion correction in diffusion-weighted imaging.

Main Methods:

• A U-net architecture, previously used for optic flow estimation, was adapted and trained to predict distortion maps.
• The network was trained by minimizing a loss function based on corrected image pairs, avoiding the need for ground truth distortion maps.
• The deep learning model was trained on data from 22 healthy subjects and tested on 12 patients with diverse acquisition modes.

Main Results:

• The deep convolutional network achieved correction accuracy comparable to iterative methods like FSL's topup.
• The developed method demonstrated a substantial reduction in computational time, processing volumes in seconds versus minutes.
• The approach proved effective even with unseen acquisition modes and patient data.

Conclusions:

• Deep convolutional networks offer a rapid and accurate solution for EPI geometric distortion correction.
• This method significantly accelerates processing, making real-time distortion correction feasible.
• The findings support the integration of this deep learning approach for all diffusion-weighted acquisitions, enhancing data quality and usability.