LAUSR

M ultiple sclerosis (MS) is a chronic demyelinating disorder of the central nervous system that mainly affects young adults. The estimated prevalence of MS in North America in 2016 was more than 120 cases per 100,000 people, and its prevalence has substantially increased worldwide from 1990 to 2016. Magnetic resonance imaging (MRI) serves an important role in the diagnosis and surveillance of MS through the detection and follow-up of focal and diffuse lesions. Recently, there has been increased attention in the cortical involvement of MS. It has been revealed that cortical lesions are closely associatedwith cognitive deficits in patients withMS and contribute to cognitive impairment irrespective of white matter lesions. Furthermore, the detection of cortical and juxtacortical lesions may contribute to early diagnosis because these lesions are relatively specific toMS. Even though cortical and juxtacortical lesions are common in MS and are included in the latest diagnostic criteria, the sensitivity of conventional MRI, such as T2-weighted and fluid-attenuated inversion recovery (FLAIR) imaging, is limited. Specialized MRI sequences, such as double inversion recovery (DIR) and phase-sensitive inversion recovery, are required to reliably visualize cortical lesions. Although DIR has been suggested to be accelerated by compressed sensing, DIR is not widely used in clinical practice because of its additional scan time. Because DIR is assumed to share information regarding T1 and T2 relaxation times and proton density with conventional contrast-weighted images, creating DIR from a combination of conventional MRI data is feasible. This could be performed by image domain transfer using deep learning, where images are created from a different domain (eg, contrast-enhanced T1-weighted image from non–contrastenhanced MRI and magnetic resonance angiography from relaxometry source data). Generative adversarial networks (GAN) have recently attracted a lot of attention in medical imaging for performing image domain transfer. The GAN algorithm uses an image generator, which generates a new image similar to the target from the input image, and a discriminator, which differentiates the target and generated images. The image generator and discriminator are simultaneously trained to transcend each other; this process is called adversarial training. In this issue of Investigative Radiology, Finck et al investigated the utility of synthetic DIR created with DiamondGAN in the detection of MS focal lesions. DiamondGAN is a GAN model with a diamond-shaped topology, where multiple input images are synergistically utilized for generating output images. A pair of generators and discriminators is used in this approach, where generators 1 and 2 simultaneously learn the mapping functions from the input to target and target to input images, respectively. Discriminator 1 discriminates the real images in the input domain and synthetic images from Generator 2, whereas discriminator 2 discriminates the real images in the target domain and synthetic images created from generator 1 (see Fig. 1 in the article by Finck et al outlining DiamondGAN). In this adversarial training process, the 4 networks are simultaneously optimized to generate images in the target domain. Notably, in this model, the input and target images do not have to be strictly spatially aligned, so the effect of misregistration due to conventional registration methods can be avoided. The authors' prior work describing the detailed technical background showed that the overall image quality of true and synthetic DIR is similar to each other, and 14 neuroradiologists were unable to distinguish between these images. In the current study, Finck et al compared synthetic DIR, true DIR, and FLAIR images in the detection of MS focal lesions. Training was performed on the data of 50 patients with MS scanned on a 3-T scanner, by using T1-weighted, T2-weighted, and FLAIR images as the input and the true DIR images as the target. The trained model evaluated another cohort of 50 patients with MS imaged on the same scanner by 2 neuroradiologists, both with 3 years of experience. As a result, the contrast-to-noise ratio of