頭部 MRI 領域における深層学習のためのモーションアーチファクトジェネレータの開発

抄録

Purpose: We focused on deep learning for a reduction of motion artifacts in MRI. It is difficult to collect a large number of images with and without motion artifacts from clinical images. The purpose of this study was to create motion artifact images in MRI by simulation. Methods: We created motion artifact images by computer simulation. First, 20 different types of vertical pixel-shifted images were created with different shifts, and the amount of pixel shift was set from –10 to 10 pixels. The same method was used to create pixel-shifted images for horizontal shift, diagonal shift, and rotational shift, and a total of 80 types of pixel-shifted images were prepared. These images were Fourier transformed to create 80 types of k-space data. Then, phase encodings in these k-space data were randomly sampled and Fourier transformed to create artifact images. The reproducibility of the simulation images was verified using the deep learning network model of U-net. In this study, the evaluation indices used were the structural similarity index measure (SSIM) and peak signal-to-noise ratio (PSNR). Results: The average SSIM and PSNR for the simulation images were 0.95 and 31.5, respectively; those for the clinical images were 0.96 and 31.1, respectively. Conclusion: Our simulation method enables us to create a large number of artifact images in a short time, equivalent to clinical artifact images.