BrainGAN Framework for Generating Synthetic Contrast-Enhanced Multisequence MRI in Brain Metastases

Contrast-enhanced MRI (CE-MRI) is routinely used to evaluate brain metastases (BMs); however, concerns regarding gadolinium-based contrast agents have prompted interest in alternative image generation strategies. This study investigated the technical feasibility of generating synthetic contrast-enhanced T1-weighted (CE-T1) and FLAIR (CE-FLAIR) images from noncontrast T1- and T2-weighted MRI using the BrainGAN generative adversarial network (GAN) framework. This retrospective study included 83 patients with 16,250 matched T1/CE-T1 pairs and 100 patients with 2810 matched T2/CE-FLAIR pairs. Within the BrainGAN framework, four GAN architectures (Pix2PixHD, CycleGAN, C-CycleGAN, and CGAN) were trained using boundary-aware conditional inputs. Data were split into training (70%), validation (20%), and testing (10%) sets. Image synthesis performance was evaluated using adversarial and VGG-based perceptual losses and quantitative metrics including mean squared error (MSE), structural similarity index measure (SSIM), peak signal-to-noise ratio (PSNR), and root mean squared error (RMSE). Qualitative evaluation employed a visual Turing test conducted by expert neuroradiologists and repeated after 1 month. GAN losses decreased progressively, and VGG-based perceptual losses from T1- and T2-weighted inputs showed strong correlation (r = 0.90). Pix2PixHD achieved the highest performance (T1: SSIM 0.90, PSNR 29.2 dB; T2: SSIM 0.90, PSNR 27.3 dB), while CGAN showed the lowest (T1: SSIM 0.30, PSNR 19.1 dB; T2: SSIM 0.20, PSNR 18.5 dB). The overall visual Turing test accuracy was 61.4%. The BrainGAN framework enables the synthesis of CE-T1 and CE-FLAIR images from noncontrast MRI with consistent quantitative performance and reproducible qualitative characteristics, supporting the technical feasibility of multisequence synthetic contrast generation in neuro-oncologic imaging.