REconstruction of MR images acquired in highly inhOmogeneous fields using Deep Learning (REMODEL)

REconstruction of MR images acquired in highly inhOmogeneous fields using Deep Learning                            (REMODEL)
 

 Asha K. Kumara Swamy1, Punith B. Venkate Gowda1 , Sachin Jambawalikar2 , Thomas Vaughan2 and Sairam Geethanath1,2*

 1Medical Imaging Research Centre, Dayananda Sagar College of Engineering, Bangalore, India,

2Dept. of Radiology, Columbia University Medical Center,New York, NY, United States

Purpose

  • To develop and demonstrate a supervised learning reconstruction method to enable imaging in highly inhomogeneous fields (or off-resonance artifacts of ±50kHz).

Introduction

  • Homogeneous magnetic field (<5ppm) is a stringent requirement  to achieve clinically acceptable  MR images.
  • Which is increasing the expense of the superconducting magnet and hence the price of the MR machine.
  • Reduction and/or relaxation of this constraint would have a significant impact on cost and access, resulting in increased MR value.
  •  The REMODEL method allows for reconstruction of MR images with comparable image quality to Ground Truth (GT), in the presence of high off resonance artifacts.

Methods

  •  Sagittal T1 weighted MPRAGE images from 70 subjects were collected from the Human Connectome Project (HCP) for training the neural network.
  •  It is then augmented by converting to axial and coronal images(using MeVis Lab, Fraunhofer MeVis, Germany) and also these images were rotated in 90o increments.
  •  Off resonance artifacts were introduced by generating 12 random field maps.(Fig b)
  • The training data was prepared by multiplying the kspace of the image with one of twelve randomly selected modulation phase maps (figure 1a).

 The neural network used for training has 3 fully connected layers, 3 convolution layers followed by one deconvolution layer, similar to the network in ref.2. Tensorflow (Google Inc, USA) was the computational framework employed to train this neural network. The kspace (nxn complex) was then reshaped to 2n x1 real valued vector and given as input to first fully connected layer (FC1). The architecture is detailed in figure 1c. All layers were activated by Relu activation function.

Model architecture:

The neural network used for training has 3 fully connected layers, 3 convolution layers followed by one deconvolution layer, similar to the network in ref.2. Tensorflow (Google Inc, USA) was the computational framework employed to train this neural network. The kspace (nxn complex) was then reshaped to 2n x1 real valued vector and given as input to first fully connected layer (FC1). The architecture is detailed in figure 1c. All layers were activated by Relu activation function.

Figure1: a) Off resonance inputs of ±50kHz used for training b) flow chart of the REMODEL method c) REMODEL’s deep neural network architecture

 

Training details:

  • trained the network using 32000 images of size 32x32,64x64 and 96x96

     with an off-resonance frequency range of ±10kHz, and ±30k and ±50kHz

     respectively. 

  •  Xavier initialization was used to initialize weights and applied mini batch normalization across layers.
  •  batch size used for training was 128
  • Learning rate of 10-4
  •  Adam optimizer was used for training
  • The loss function used for training was mean squared loss between the network output and target image intensity values.
  •  used an additional L2 norm and regularization value of λ=10-3.
  • Each network was trained for 600 epochs using 2 G-Force GTX 1080 Ti GPUs with 11GB memory capacity each.

Validation:

  •  The performance of the resulting network was tested using the images downloaded from HCP that were not included for training and corrupted with random field maps.

Results

 Figure 2 shows the 96x96 images corrupted by ±50kHz off-resonance.

  • The figure shows the blurring and pixel shifting due to inhomogeneity. It can be observed that these artifacts are significantly removed by REMODEL.
  •  The line intensity profiles are similar to those shown for GT. Analogous results can be seen in axial and sagittal orientations in figure 2.
  • The average and standard deviation of the Root Mean Square Error plot of over 190 examples in figure 3 quantitates the reduction of artifacts by REMODEL reconstruction with reference to the GT, as compared to the corrupted images.

 

Figure 2: REMODEL applied on corrupted brain data in three orientations  along with representative line intensity profiles

 

Figure 3: Root Mean Square Error plots over 190 examples for REMODEL compared to GT and corrupted data

 

 

Discussion and Conclusion

 REMODEL has been demonstrated on synthetically corrupted data from a publicly available database. It is able to faithfully reconstruct the images simulated in inhomogeneous fields up to ±50kHz, which can be extended to higher off-resonance ranges and to other body regions and for multiple readout times without any change in implementation details. REMODEL is contrast and trajectory dependent. Implementation details along with source code can be found online[3].

Current & Future work :REMODEL has been trained for EPI trajectory with off resonance range of ±1000Hz and the results are shown in figure 4. It can be observed that REMODEL is able to correct these artifacts with blurring. This was attributed to a kernel size of 3 x 3 for an image of size
96x96. However, this can be reduced by tuning suitable filter size for convolution and deconvolution operation or applied on an image with larger matrix size.

 Figure 4: REMODEL applied on the image corrupted by EPI trajectory and off resonace range of ±1000Hz

References

 (1) Bernstein et. al. Handbook of MRI pulse sequences 2004. (2) Zhu, Bo et. al. arXiv 2017 (3) https://github.com/mirc-dsi/IMRI REMODEL_v0.0