Elsevier

Magnetic Resonance Imaging

Volume 71, September 2020, Pages 1-10
Magnetic Resonance Imaging

Original contribution
An unsupervised deep learning technique for susceptibility artifact correction in reversed phase-encoding EPI images

https://doi.org/10.1016/j.mri.2020.04.004Get rights and content

Highlights

  • We propose a deep learning technique, named S-Net, to correct the susceptibility artifacts in a reversed-PE EPI image pair.

  • The proposed S-Net consists of two components: a convolutional neural network (CNN) and a spatial transform unit (STU).

  • The CNN is to map a reversed-PE image pair to the displacement field, and the STU is to produce the corrected images.

  • S-Net is trained in an end-to-end manner on a set of reversed-PE image pairs using an unsupervised learning algorithm.

  • Compared to two other state-of-the-art methods, the proposed S-Net runs significantly faster while achieving similar accuracy.

  • The fast processing speed of the S-Net technique enables performing susceptibility artifact correction in MRI scanner.

Abstract

Echo planar imaging (EPI) is a fast and non-invasive magnetic resonance imaging technique that supports data acquisition at high spatial and temporal resolutions. However, susceptibility artifacts, which cause the misalignment to the underlying structural image, are unavoidable distortions in EPI. Traditional susceptibility artifact correction (SAC) methods estimate the displacement field by optimizing an objective function that involves one or more pairs of reversed phase-encoding (PE) images. The estimated displacement field is then used to unwarp the distorted images and produce the corrected images. Since this conventional approach is time-consuming, we propose an end-to-end deep learning technique, named S-Net, to correct the susceptibility artifacts the reversed-PE image pair. The proposed S-Net consists of two components: (i) a convolutional neural network to map a reversed-PE image pair to the displacement field; and (ii) a spatial transform unit to unwarp the input images and produce the corrected images. The S-Net is trained using a set of reversed-PE image pairs and an unsupervised loss function, without ground-truth data. For a new image pair of reversed-PE images, the displacement field and corrected images are obtained simultaneously by evaluating the trained S-Net directly. Evaluations on three different datasets demonstrate that S-Net can correct the susceptibility artifacts in the reversed-PE images. Compared with two state-of-the-art SAC methods (TOPUP and TISAC), the proposed S-Net runs significantly faster: 20 times faster than TISAC and 369 times faster than TOPUP, while achieving a similar correction accuracy. Consequently, S-Net accelerates the medical image processing pipelines and makes the real-time correction for MRI scanners feasible. Our proposed technique also opens up a new direction in learning-based SAC.

Introduction

Echo Planar Imaging (EPI) is the technique of choice for most functional magnetic resonance imaging (fMRI) and diffusion-weighted imaging (DWI) applications due to its fast imaging capability. Despite its popularity, EPI is sensitive to the local field inhomogeneities, which are caused by magnetic susceptibility differences of various imaged tissues, e.g. air versus bone, and fat versus blood [1,2]. The field inhomogeneities affect the spatial encoding of the signal, resulting in intensity modulations and local image distortions (i.e. stretching and compressing) [3]. These distortions are known as susceptibility artifacts (SAs). They cause the misalignment to the underlying structural image, subsequently leading to incorrect localization of analysis results, such as the wrong activation patterns in the fMRI studies. SAs are more severe at high field strengths [4,5], which have become widely used.

Several susceptibility artifact correction (SAC) methods rely on two reversed phase-encoding (PE) images, which are acquired using identical sequences but with opposite PE directions. The main idea is that the SAs appear inversely in the pair of reversed-PE images [[6], [7], [8]]; therefore, the middle version of the reversed-PE image pair is considered the corrected image [3]. Conventionally, the reversed-PE SAC methods involve two steps. In the first step, a displacement field along the PE direction is estimated by optimizing an objective function of one or more reversed-PE image pairs. Only the displacement field1 along the PE direction is estimated since the displacements in the other directions are less significant [2,9,10]. In the second step, the estimated displacement field is used to unwarp the distorted images and produce the corrected images. This conventional approach is time-consuming, especially for input images with large sizes or severe displacements. Consequently, these SAC methods are unsuitable for time-sensitive applications, for example correction on an MRI scanner.

To reduce the processing time, we propose an unsupervised deep learning technique, called S-Net, for correcting the susceptibility artifacts in 3D reversed-PE images. A convolutional neural network (CNN) is used to map a pair of reversed-PE images to the displacement field in the PE direction. Then, a differentiable spatial transform unit is used to unwarp the input (distorted) image pair via the predicted displacement field. The S-Net is trained in an end-to-end manner using a training set of reversed-PE image pairs. After training, correcting a new 3D reversed-PE image pair is achieved by simply evaluating the trained S-Net on the given input images. This approach, therefore, avoids the highly computational cost of the existing iterative optimization approaches.

The contributions of this paper are highlighted as follows.

  • 1.

    We design a convolutional encoder-decoder network to map a 3D reversed-PE image pair to the displacement field. The network consists of an encoder for image downsampling and a decoder for image upsampling. The encoder uses a series of convolutional (conv) layers and leaky rectified linear units (LeakyReLUs) to extract hierarchical image contents. The decoder uses a series of conv layers, LeakyReLUs, and upsampling layers to recover the full-resolution image features and estimate the displacement field. A spatial transform unit is designed to unwarp the 3D input images along the PE direction and produce an output image pair. To the best of our knowledge, the proposed technique is the first attempt at using a learning-based approach with an convolutional encoder-decoder to correct the susceptibility artifacts.

  • 2.

    We explore an unsupervised learning strategy in design the proposed S-Net. The term unsupervised learning arises from the fact that the S-Net is trained without additional ground-truth information, e.g. the “desired” displacement field or the “desired” corrected images, which are impractical to acquire. In our approach, the S-Net is trained by maximizing the similarity of the output image pairs and the smoothness of the displacement field, which are inspired by the traditional reversed-PE based SAC methods.

  • 3.

    We evaluate the performance of the proposed technique and compare it with existing SAC methods using three datasets. The datasets include one dataset acquired by our team using a 7 Tesla (T) scanner, and two public datasets acquired using a 3 T scanner and published by the Human Connectome Project (HCP) [11]. The experimental results show that our unsupervised S-Net provides the corrected images which are comparable to results of state-of-the-art SAC methods while requiring fewer computational resources and no additional data, such as structural images.

The remainder of this paper is organized as follows. Section 2 presents the related work, and Section 3 introduces the proposed method. Section 4 presents experiments and analysis of the proposed method and the related methods. Finally, Section 5 summarizes our work.

Section snippets

Related work

In this study, we investigate the reserved-PE based SAC because of its efficiency and popularity. This SAC approach is used to correct the fMRI and DWI data in the biggest MRI neuroimaging dataset - the HCP with 1200 subjects of multiple MRI modalities [12,13]. Recall that the reserved-PE SAC first estimates the displacement field based on a pair of images acquired using an identical sequence but with opposite PE directions. The corrected images are then obtained by unwarping the distorted

The proposed deep learning technique for SAC

This paper introduces a deep learning technique, called S-Net, for correcting a pair of 3D reversed-PE images. Fig. 1 illustrates an overview of the proposed S-Net. It consists of two parts: (i) a mapping function to estimate the displacement field U from a pair of 3D reversed-PE images I1 and I2; and (ii) a spatial transform unit to recover the corrected images by unwarping the input images with the estimated 3D displacement field. Note that the computations in the S-Net are performed in 3D.

Experiments and analysis

This section presents the experiments and analysis of the proposed S-Net. Section 4.1 describes the datasets used in the experiments, and Section 4.2 explains the experimental methods. Section 4.3 analyzes the regularization and learning rate parameters. 4.4 Comparison with other methods in correction accuracy, 4.5 Comparison with other methods in processing speed compare the correction accuracy and processing time of the proposed method with other representative methods, respectively. Finally,

Conclusion

This paper introduced a novel unsupervised deep learning technique, S-Net, for correcting susceptibility artifacts in reversed-PE EPI images in an end-to-end setting. The proposed S-Net contains a convolutional encoder-decoder to map a reversed-PE image pair to the displacement field. The displacement field is then fed to spatial transform units to unwarp the input images, resulting in the corrected images. The S-Net is trained in an unsupervised manner, without requiring the ground-truth data.

CRediT authorship contribution statement

Soan T.M. Duong: Conceptualization, Investigation, Data curation, Methodology, Software, Formal analysis, Validation, Visualization. Son L. Phung: Conceptualization, Methodology, Resources, Supervision, Funding acquisition. Abdesselam Bouzerdoum: Funding acquisition, Writing - review & editing. Mark M. Schira: Conceptualization, Methodology, Resources, Supervision, Funding acquisition.

Declaration of competing interest

The authors have declared that no competing interests exist.

Acknowledgment

The authors acknowledge the National Imaging Facility at the Center for Advanced Imaging, University of Queensland. The authors also thank Siemens Healthineers for providing the prototype WIP1080. This research was supported by grants (DP170101778 and DP190100607) from the Australian Research Council and a Matching scholarship from the University of Wollongong.

References (43)

  • K. Uğurbil et al.

    Pushing spatial and temporal resolution for functional and diffusion MRI in the Human Connectome Project

    NeuroImage

    (2013)
  • H. Chang et al.

    A technique for accurate magnetic resonance imaging in the presence of field inhomogeneities

    IEEE Trans Image Process

    (1992)
  • S. Ogawa et al.

    Brain magnetic resonance imaging with contrast dependent on blood oxygenation

  • P. Jezzard et al.

    Correction for geometric distortion in echo planar images from B0 field variations

    Magn Reson Med

    (1995)
  • D. Holland et al.

    Efficient correction of inhomogeneous static magnetic field-induced distortion in echo planar imaging

    NeuroImage

    (2010)
  • L. Ruthotto et al.

    Diffeomorphic susceptibility artifact correction of diffusion-weighted magnetic resonance images

    Phys Med Biol

    (2012)
  • D.C.V. Essen et al.

    The human connectome project: a data acquisition perspective

    NeuroImage

    (2012)
  • D.C.V. Essen et al.

    The WU-Minn human connectome project: an overview

    NeuroImage

    (2013)
  • S.T.M. Duong et al.

    Anatomy-guided inverse-gradient susceptibility artefact correction method for high-resolution fMRI

  • S.T.M. Duong et al.

    Susceptibility artifact correction for sub-millimeter fMRI using inverse phase encoding registration and T1 weighted regularization

    J Neurosci Methods

    (2020)
  • R. Bowtell et al.

    Correction of geometric distortion in echo planar images

    Proc Int Soc Magn Reson

    (1994)
  • Cited by (21)

    • Potential of Deep Learning in Quantitative Magnetic Resonance Imaging for Personalized Radiotherapy

      2022, Seminars in Radiation Oncology
      Citation Excerpt :

      An interesting example of denoising the raw images was done by training CNNs to turn single average images into images from multiple averages.90 Furthermore, deep learning can be used to reduce artefacts.91 It was also shown that deep learning can be used to optimize acquisition settings.92

    • EPI susceptibility correction introduces significant differences far from local areas of high distortion

      2022, Magnetic Resonance Imaging
      Citation Excerpt :

      This indirect structural-based transformation procedure is often the preferred method for normalizing EPI images since computing a deformation from a structural image and a template is generally easier, due to the high contrast in high-resolution structural images, than computing a deformation between a subject's EPI image and a template. Unfortunately, EPI distortion introduces geometric deformations between the EPI image and high-resolution structural image, though reducing EPI distortion to recover brain geometry has been shown to raise the accuracy of registration between DWIs and high-resolution structural images [12–20]. The influence of correcting for EPI distortion on inference during diffusion-based applications had been noted in several studies; due to the increased correspondence in in the anatomical overlap between subjects, EPI distortion correction has been shown to enhance statistical power in group studies [12], and improve re-test reliability of group comparisons involving diffusion analysis [21], including regions of the brain relatively remote from source of EPI distortion [22].

    View all citing articles on Scopus
    View full text