Improved RED-CNN with Channel-Spatial Attention and Hybrid Loss for Low-Dose CT

In low-dose medical CT imaging, reduced radiation exposure often leads to streak artifacts, detail loss, and decreased resolution, thereby compromising diagnostic accuracy. To address these unique challenges, the proposed model incorporates a channel-spatial attention mechanism (CSA) and a structure-aware composite loss function into the RED-CNN framework. The CSA module was designed to jointly model inter-channel dependencies and spatial contextual information, effectively highlighting critical regional features. Additionally, a structural similarity index was incorporated into the loss function to improve structural fidelity while maintaining pixel-level accuracy. These design choices directly target the degradation patterns typical of low-dose CT scans, such as noise and structural blurring. Experiments conducted on the Mayo datasets demonstrated that the proposed method outperforms several existing denoising techniques in terms of PSNR and SSIM, indicating its strong potential for clinical application.