This study used brain MR images for screening or workup of mild cognitive impairment (MCI)/dementia and epilepsy, including MPRAGE, taken at a single institution between January 2020 and December 2022. In this study, patient selection criteria for MCI/dementia meet the clinical diagnostic criteria for probable Alzheimer's disease (AD) dementia and MCI due to AD based on the 2011 guidelines of the National Institute on Aging and the Alzheimer's Association. Clinical diagnosis was performed by board-certified neurologists. Regarding epilepsy, patients with epilepsy diagnosed by board-certified neurologists underwent an “epilepsy MRI protocol” including MPRAGE [20]. In the other MRI protocols, MPRAGE was not used at our institution due to MRI examination slots. We excluded cases showing excessive motion artifacts or those in which the default resolution of MPRAGE had been changed to adjust skull size. For the test dataset, patients included in the training and validation datasets were excluded to avoid affecting evaluations of the test dataset. Images taken between 2020 and 2021 were assigned to training and validation datasets at a ratio of 9:1. All images from 2022 were used for the test dataset because it was difficult to use datasets from other institutes or public datasets due to the use of localizer images and the need to increase the number of cases to verify clinical utility. This retrospective study was approved by the institutional ethics committee. All MR images had only been taken out of clinical necessity, and the need to obtain written informed consent was waived based on the retrospective nature of the work. No subjects overlapped with previously published work.

Acquisitions were performed using 3.0-T whole-body systems (Magnetom Skyra, Prisma, and Vida; Siemens Healthineers, Erlangen, Germany) using 32-, 64-, and 32-channel receive-only head coils, respectively. Imaging parameters were as follows. AAH (3D FLASH): TR, 3.15 ms; TE, 1.37 ms; flip angle (FA), 8°; bandwidth, 540 Hz/pixel; spatial resolution, isotropic voxels of 1.6 mm; and slices, 128. For acceleration, 24 reference lines were acquired in the phase-encoding direction, and generalized autocalibrating partially parallel acquisition (GRAPPA) 3 × was used. The acquisition time was 14 s. MPRAGE: TR, 2300 ms; TE, 4.67 ms; FA, 9°; bandwidth, 130 Hz/pixel; spatial resolution, isotropic voxels of 0.9 mm, and slices, 208. For acceleration, 24 reference lines were acquired in the phase-encoding direction, and GRAPPA 2 × was used. The acquisition time was 4 min 26 s.

We used the pix2pix method, which combines a U-Net generator with a conditional GAN [10]. Pix2pix is widely used for image-to-image translation and has reportedly been used for medical images [18, 21, 22]. We modified the pix2pix code in pytorch (https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix) [10]. Details of the model modification and training procedure are described in the Supplementary material. The code used in this study is available on GitHub (https://github.com/kuponuga/aah2mprage).

In training, validation, and testing, we used peak signal-to-noise ratio (PSNR, with higher values considered better), structural similarity index measure (SSIM, with higher values considered better) [23], and Learned Perceptual Image Patch Similarity (LPIPS, with lower values considered better) [24] (https://github.com/richzhang/PerceptualSimilarity) as metrics for assessing image quality. The default Alex network was used in LPIPS.

Two radiologists (A.S. and H.T., with 15 and 10 years of experience in neuroradiology, respectively) performed visual evaluation of generated images and determined the best model. First, we extracted eight slices that included the brain, reviewed the images for each condition, and excluded those that showed inaccurate translations or apparent artifacts. We then chose those with superior image quality based on visual evaluation and finally selected the best model based on a consensus of two radiologists. For the selected model, we looked through all the validation images to check for any abnormalities that would cause significant problems for evaluation. Image evaluation indices (PSNR, SSIM, LPIPS) were also calculated to see if they differed from the radiologists’ evaluations.

PSNR, SSIM, and LPIPS were calculated between the original MPRAGE and generated images from the test datasets only for those sections containing brain parenchyma. As with the training procedure (Supplementary material), one radiologist (H.T.) excluded those slices not containing brain parenchyma.

VBM analysis was performed using FreeSurfer (version 7.4.1, https://surfer.nmr.mgh.harvard.edu/fswiki/FreeSurferWiki). First, image processing and segmentation were performed on the original MPRAGE and on generated images using the FreeSurfer “recon-all” command. Considering the influence of manual correction on analysis results, we excluded images from evaluation without manual correction if an error was identified. Concordance of the segmentation regions of major structures (each side of the thalamus, caudate, putamen, hippocampus, amygdala, cerebral white matter, cerebral cortex, and lateral ventricle, third ventricle, and fourth ventricle) was evaluated using Dice scores. These Dice scores were calculated using the “mri_overlap” command on FreeSurfer. We also evaluated volume differences as the absolute symmetrized percent change (ASPC). We used the volumes provided from FreeSurfer recon-all stats. Dice score and ASPC are defined as follows:

$$Dice\left(X, Y\right)= \frac$$

$$ASPC\left(X, Y\right)=\frac$$

Three radiologists (S.Ik., S.It., and M.U., each with 7 years of experience in neuroradiology) evaluated both MPRAGE and generated images of the test dataset obtained in 2022. The presence of medial temporal lobe atrophy (MTA) and old cerebral infarction or hemorrhage was visually assessed. The following medial temporal lobe atrophy score [25] was used to evaluate atrophy: 0, normal; 1, widened choroid fissure; 2, increased widening of the choroid fissure, widening of the temporal horn, opening of other sulci (i.e., collateral/fusiform sulcus); 3, pronounced loss of hippocampal volume; and 4, end-stage atrophy. Cerebrovascular lesions were defined as those with a short diameter ≥ 1 cm. Generated images were evaluated first. Four weeks later, MPRAGE images were evaluated next. Raters were not informed which images were the generated images and which were MPRAGE, and the order of cases was randomized. For lesions not detected on generated images, another radiologist (H.T.) reviewed the images. Conspicuous artifacts were also recorded. For MTA scores, weighted Cohen kappa was calculated between MPRAGE and the generated images. Quadratic weighting was used to emphasize the large difference in scores. Statistical analysis was performed with R (version 4.3.1, https://www.r-project.org/) on RStudio (version 2023.09.1, https://posit.co/download/rstudio-desktop/).