Fig 1.
Schematic diagram of this study.
Artifact images were created by manipulating the k-space of the original images. Motion artifact suppressed images were generated from the artifact images using the Pix2Pix network. VSRAD analysis was performed on three types of images: the original image, the artifact image, and the motion-corrected image. Then, differences among images of the VSRAD analysis results were examined.
Fig 2.
Creation of motion artifact images.
The following process was used to create motion artifact images. First, two images were created, one rotated to the left and one rotated to the right of the image. Secondly, the two types of images were Fourier transformed (FFT), and we filled k-space with zeros at equal intervals. Finally, the top and bottom half of each k-space were merged to create one image, and we performed an inverse Fourier transform (IFT) of the merged image to create an image with simulated motion artifacts.
Fig 3.
A total of 16 artifact images were created with four rotation angles (RA) and four patterns of k-space zero-fill intervals (KI). In the artifact images, the larger the rotation angle, or the smaller the zero-fill interval in k-space, the stronger the motion artifact. In the present study, we created a training data set by pairing each of the 16 artifact images with the original image. We demonstrate an example used as the input image for the training group.
Fig 4.
We used a U-Net-based network for the generator network. This network comprised an encoder consisting of eight convolutional layers and a decoder consisting of the same number of deconvolutional layers. For the convolutional layer’s input and output, activation by the LeakyReLu function and batch normalization (BN) were performed in the encoder stack, whereas activation by the ReLu function and BN was performed in the decoder stack. The motion-corrected image was the output of the decoder stack activated by the tanh function.
Fig 5.
PachGAN was used as the Discriminator network. First, the input image was cropped into small regions (patch size). Secondly, for each small region, binary classification was performed by a network consisting of a convolutional layer, an activation layer (LeakyReLu, Softmax), a batch normalization layer (BN), and an affine layer (Flat). Thirdly, the outputs of the binary classification in all subregions were combined. Finally, binary classification was performed using the Softmax function to evaluate whether the two input images were entirely equal.
Table 1.
Details of the training and test groups.
Fig 6.
Examples of the original image, artifact image, and motion-corrected image.
(A) The original T1-weighted, magnetization-prepared rapid gradient-echo sequence (T1-MPRAGE) images, (B) artifact images created by manipulating the k-space of T1-MPRAGE, (C) motion-corrected images by the U-Net network from the artifact images and (D) motion-corrected images by the Pix2Pix network from the artifact images, are shown, respectively. The rotation angle and k-space interval of the artifact images were 2°_20 pixels.
Fig 7.
Comparison of SSIMs in the original and artifact images or generated images by the U-net or Pix2Pix.
The measured SSIM values are shown in a box-and-whisker diagram. A comparison of the artifact image vs. the original images are shown in black, the motion-corrected image created by U-Net vs. the original images are shown in red, andPix2Pix-corrected vs. the original images are shown in blue. The SSIMs of the motion-corrected images were compared by paired t-test with the same rotation angle and k-space interval, and Pix2Pix was significantly higher than U-Net (p < 0.05).
Table 2.
Results of SSIM analysis and qualitative evaluation when set to trained rotation angles and k-space intervals.
Fig 8.
The Bland-Altman plot for VSRAD analysis results.
Bland-Altman analysis results are shown for (A): severity of voxel of interest (VOI) atrophy of artifact vs. the original image, (B): severity of voxel of interest (VOI) atrophy of U-Net motion-corrected vs. the original image, (C): severity of voxel of interest (VOI) atrophy of Pix2Pix motion-corrected vs. the original image. The dotted line and white boxes plot show the relationship between the ±1.96 standard deviation(SD) limit of agreement of the artifact image and the original image, and the dashed line show the relationship between the ±1.96 SD limit of agreement of the motion-corrected image for U-Net or Pix2Pix and the original image. The rotation angle and k-space interval are plotted as black circles for 1°_30 pixel, red squares for 2°_20 pixel, and blue triangles for 3°_10 pixel.
Table 3.
Results of Spearman’s rank correlation coefficient and Bland-Altman analysis when set to trained rotation angles and k-space intervals.
Table 4.
Results of Spearman’s rank correlation coefficient analysis and Bland-Altman analysis when set to non-trained rotation angles and k-space intervals.
Fig 9.
The Spearman’s rank correlation coefficient for the results of the VSRAD analysis.
The results of the Spearman’s rank correlation coefficient are shown for (A): severity of voxel of interest (VOI) atrophy of artifact vs. the original image, (B): severity of voxel of interest (VOI) atrophy of U-Net motion-corrected vs. the original image and, (C): severity of voxel of interest (VOI) atrophy of Pix2Pix motion-corrected vs. the original image. The dotted line and white points plot show the relationship between the limits of agreement of the artifact image and the original image, and the solid line show the relationship between the limits of agreement of the motion-corrected image for U-Net or Pix2Pix and the original image. The rotation angle and k-space interval are plotted with 1°_30 pixel in black, 2°_20 pixel in red, and 3°_10 pixel in blue.