An unsupervised deep learning technique for susceptibility artifact correction in reversed phase-encoding EPI images.
Soan T M Duong1, Son L Phung1, Abdesselam Bouzerdoum2
1School of Electrical, Computer and Telecommunications Engineering, University of Wollongong, Australia.
Magnetic Resonance Imaging
|May 15, 2020
Summary
We developed S-Net, a deep learning method to rapidly correct echo planar imaging susceptibility artifacts. S-Net significantly speeds up magnetic resonance imaging processing, enabling real-time artifact correction.
Area of Science:
- Medical Imaging
- Neuroimaging
- Machine Learning
Background:
- Echo planar imaging (EPI) offers fast, non-invasive data acquisition but suffers from susceptibility artifacts causing image misalignment.
- Traditional susceptibility artifact correction (SAC) methods are time-consuming, relying on optimizing objective functions with reversed phase-encoding (PE) image pairs.
Purpose of the Study:
- To introduce S-Net, an end-to-end deep learning technique for rapid and accurate susceptibility artifact correction in EPI.
- To enable real-time artifact correction for magnetic resonance imaging (MRI) scanners.
Main Methods:
- Developed S-Net, comprising a convolutional neural network to predict displacement fields from reversed-PE image pairs and a spatial transform unit for unwarping.
- Trained S-Net using an unsupervised loss function without requiring ground-truth data.
Main Results:
- S-Net effectively corrects susceptibility artifacts in reversed-PE images across diverse datasets.
- Achieved comparable correction accuracy to state-of-the-art methods (TOPUP, TISAC) but with significantly faster processing speeds (20x faster than TISAC, 369x faster than TOPUP).
Conclusions:
- S-Net offers a substantial acceleration of medical image processing pipelines.
- The proposed deep learning approach facilitates real-time artifact correction, making it feasible for clinical MRI scanners and opening new avenues for learning-based SAC.
