Enhancing Multi-Contrast MRI Synthesis: A Novel 3D Dual-CycleGAN Approach

Abstract

Abstract This study introduces an innovative approach to synthesizing 3D Multi-Contrast MRI images utilizing the 3D Dual-CycleGAN model. The model's performance is assessed using MRI data from the Multi-modal Brain Tumour Segmentation Challenge (BraTS) 2021 dataset, encompassing T1W, T1c, T2W, and FLAIR sequences. The 3D Dual-CycleGAN model effectively learns the transformation between T1W images and the desired target modalities, leading to enhanced fine details and overall quality of the synthesized images. Incorporating advanced techniques and a comprehensive objective function, the proposed method encompasses voxel-wise, gradient difference, perceptual, and structural similarity losses. These loss functions, combined with adversarial and dual cycle-consistency losses, play a pivotal role in producing lifelike and visually captivating representations. For performance evaluation, a set of five metrics (MAE, PMAE, RMSE, PCC, PSNR, and SSIM) are employed to compare the synthesized images against their authentic counterparts. Empirical findings demonstrate the prowess of the 3D Dual-CycleGAN model in generating T1c images from T1W inputs, exhibiting minimal average discrepancies (MAE of 2.8±2.61) and strong similarity (SSIM of 0.82±0.28). Moreover, the synthesis of T2W and FLAIR images yields promising results, showcasing acceptable average discrepancies (MAE of 3.87±3.32 for T2W and 3.82±3.32 for FLAIR) and reasonable similarities (SSIM of 0.82±0.28 for T2W and 0.80±0.29 for FLAIR) when compared to the original images. These outcomes underscore the effectiveness of the 3D Dual-CycleGAN model in generating high-quality images. The implications of this achievement are substantial across various applications within the realm of medical imaging.