T۲-Weighted (T۲W) synthesis from brain Fluid-Attenuated-Inversion-Recovery (FLAIR) images and vice versa based on deep learning methods

Background and aims: In this study, considering the importance of T۲-Weighted (T۲W) andFluid-Attenuated-Inversion-Recovery (FLAIR) scans, we focus on generating FLAIR from T۲WMRI scans, and vice versa, based on the CycleGAN.Method: The data utilized in our study were obtained from the public data collection ‘ACRINDSC-MR-Brain’. We extracted ۵۱۰ T۲W/FLAIR paired slices from ۱۰۲ patients, which werefurther divided into ۴۱۰ pairs from ۸۱ patients for training and ۱۰۰ pairs from ۲۱ patients to evaluatesynthesis results. Each pair include the axial paired T۲W/FLAIR slices for the same patientand at the same axial depth. Of note, the testing data are held out of the training process at alltimes. Our networks take ۲D axial-plane slices of the volumes as inputs. The Cycle GenerativeAdversarial Network (CycleGAN) model was applied for the synthesis task. The proposed Cycle-GAN operates with two generators (GT۲W, GFLAIR) and two discriminators (DT۲W, DFLAIR).Given a T۲W image, GT۲W learns to generate the respective FLAIR image of the same anatomythat is indistinguishable from real FLAIR images, whereas DT۲W learns to discriminate betweensynthetic and real FLAIR images. Similarly, given a FLAIR image, GFLAIR learns to generatethe respective T۲W image of the same anatomy that is indistinguishable from real T۲W images,whereas DFLAIR learns to discriminate between synthetic and real T۲W images. To generatea T۲W MRI from a FLAIR, and vice versa, the T۲W and FLAIR values are converted to [۰, ۱]tensor. The resolutions of the FLAIR and T۲W images in our dataset are ۲۵۶× ۲۵۶ and ۵۱۲× ۵۱۲respectively. Therefore, in the first preprocessing step, FLAIR images are registered to T۲W imagesusing rigid registration to ensure that all images have a ۲۵۶ × ۲۵۶ resolution. Then, the axialT۲W/FLAIR pairs were the input of the network with ۲۵۶×۲۵۶ pixels. Performance evaluationis conducted based on the Mean Absolute Error (MAE), Mean Squared Error (MSE), and PeakSignal-to-Noise Ratio (PSNR) metrics.Results: It has been shown, via a perceptual study and in terms of quantitative assessments basedon MAE, MSE, and PSNR metrics, that CycleGAN can be used to generate visually realistic MRimages.Conclusion: The CycleGAN method can be used to generate realistic synthetic T۲W and FLAIRbrain scans, supported by both experimental qualitative and quantitative results.