SSC-NRNet: Adaptive Skip Compression for statistical, structural, and composite MRI noise reduction in a cGAN-based framework

Magnetic Resonance Image (MRI) denoising remains challenging because image degradation may arise from statistical noise, structural artifacts, and their composite combinations. While many existing denoising methods primarily address isolated statistical noise, their effectiveness often decreases under structurally distorted or mixed corruption conditions that more closely reflect practical MRI degradation. To address this limitation, we propose SSC-NRNet, a conditional generative adversarial network (cGAN)-based MRI denoising framework whose main architectural contribution is an Adaptive Skip Compression (ASC) mechanism integrated into the skip connections of a U-Net generator. Rather than relying solely on direct encoder-to-decoder feature transfer, ASC introduces a compression–expansion refinement pathway that supplements the conventional skip representation with a more compact and selective feature branch before decoder fusion. The proposed framework was evaluated on 3064 T1-weighted brain MRI images under simulated statistical noise, structural artifacts, and composite corruption settings using 5-fold cross-validation. In addition, external evaluation was performed on the MR-ART dataset containing real motion-corrupted MRI scans. Experimental results show that SSC-NRNet consistently outperforms both the Baseline U-Net GAN and the Denoising Cycle-GAN across the evaluated corruption settings, achieving SSIM values up to 0.97 and PSNR values up to 31.19 dB. On the MR-ART dataset, SSC-NRNet also demonstrates improved restoration quality relative to the baseline U-Net generator. Additional analyses show that moderate compression yields the most effective ASC configuration and reduces channel-wise redundancy in skip representations, supporting SSC-NRNet as an effective software-based MRI denoising framework without requiring hardware or acquisition modifications.