Abstract
Magnetic Resonance Imaging (MRI) is often constrained by long acquisition times. Accelerated acquisition techniques can reduce scan time but may introduce artifacts and decrease image resolution. Additionally, MRI data invariably contains noise (often Gaussian or Rician), further impacting image quality and diagnostic value. Conventional approaches often address noise and undersampling artifacts separately, leading to suboptimal image reconstruction. This work proposes a novel ensemble of CycleGAN models integrated within the Joint Sensitivity Encoding (JSENSE) framework to address these challenges in a comprehensive manner. Each CycleGAN within the ensemble is specifically trained to target distinct aspects of image degradation. This approach leverages the complementary strengths of the models to achieve a superior image reconstruction. Experiments demonstrate the effectiveness of the proposed ensemble model, outperforming conventional methods and our prior work in terms of visual quality and quantitative metrics. Significant gains were observed at higher acceleration factors and greater noise levels. The integration of CycleGANs with JSENSE yielded superior structural similarity and image clarity compared to traditional parallel MRI (pMRI) and single-model methods.