Deep flow-net for EPI distortion estimation

• Published in: NeuroImage (Orlando, Fla.) Vol. 217; p. 116886
• Main Authors: Zahneisen, Benjamin, Baeumler, Kathrin, Zaharchuk, Greg, Fleischmann, Dominik, Zeineh, Michael
• Format: Journal Article
• Language: English
• Published: United States Elsevier Inc 15.08.2020; Elsevier Limited; Elsevier
• Subjects: Adult; Algorithms; Brain Mapping; Computer applications; Computer Simulation; Databases, Factual; Diffusion Magnetic Resonance Imaging - methods; Diffusion Magnetic Resonance Imaging - statistics & numerical data; Echo-Planar Imaging - methods; Echo-Planar Imaging - statistics & numerical data; Humans; Image Processing, Computer-Assisted - methods; Machine Learning; Mapping; Neural Networks, Computer; Optic flow; Software packages
• ISSN: 1053-8119; 1095-9572; 1095-9572
• DOI: 10.1016/j.neuroimage.2020.116886

Geometric distortions along the phase encoding direction caused by off-resonant spins are a major issue in EPI based functional and diffusion imaging. The widely used blip up/down approach estimates the underlying distortion field from a pair of images with inverted phase encoding direction. Typically, iterative methods are used to find a solution to the ill-posed problem of finding the displacement field that maps up/down acquisitions onto each other. Here, we explore the use of a deep convolutional network to estimate the displacement map from a pair of input images. We trained a deep convolutional U-net architecture that was previously used to estimate optic flow between moving images to learn to predict the distortion map from an input pair of distorted EPI acquisitions. During the training step, the network minimizes a loss function (similarity metric) that is calculated from corrected input image pairs. This approach does not require the explicit knowledge of the ground truth distortion map, which is difficult to get for real life data. We used data from a total of Ntrain ​= ​22 healthy subjects to train our network. A separate dataset of Ntest ​= ​12 patients including some with abnormal findings and unseen acquisition modes, e.g. LR-encoding, coronal orientation) was reserved for testing and evaluation purposes. We compared our results to FSL’s topup function with default parameters that served as the gold standard. We found that our approach results in a correction accuracy that is virtually identical to the optimum found by an iterative search, but with reduced computational time. By using a deep convolutional network, we can reduce the processing time to a few seconds per volume, which is significantly faster than iterative approaches like FSL’s topup which takes around 10min on the same machine (but using only 1 CPU). This facilitates the use of a blip up/down scheme for all diffusion-weighted acquisitions and potential real-time EPI distortion correction without sacrificing accuracy. •Processing speed argument: We ran all methods on a single core virtual machine and still found our method to perform significantly faster than topup. However, in light of other correction methods that are optimized for speed, we modified several statements to account for that.•We added all the suggested references to the manuscript.•We included the previously missing table 1.