ARUC-GAN: A CycleGAN-Based Attention Residual U-Net for Low-Dose CT Denoising in Smart Healthcare

Low-dose computed tomography (LDCT) is increasingly adopted in medical imaging to minimize radiation exposure. However, the diagnostic accuracy is primarily affected by different noise sources, e.g., quantum, electronic, and reconstruction. Moreover, conventional denoising methods struggle with the non-uniform noise distributions in LDCT images and often require complex projection data, which limits their effectiveness and generalizability. Herein, we propose an AI-based denoising approach using an Attention Residual U-Net (ARU-Net) architecture integrated into the CycleGAN framework named Attention Residual U-Net CycleGAN (ARUC-GAN). The proposed framework outperforms state-of-the-art denoising models, such as RED-CNN, EDCNN, and CTformer, according to experimental evaluation on the Mayo Clinic abdominal CT dataset. The model exhibits a peak signal-to-noise ratio (PSNR) of 34.82 dB, a structural similarity index (SSIM) of 0.85, and an RMSE of 0.018. Furthermore, the model successfully maintains edge structures with an Edge Keeping Index (EKI) of 0.835. The visual examination validates ARUC-GAN’s superior texture and detail preservation. The results indicate that the suggested method has great promise as a smart health tool for improving diagnosis accuracy in LDCT scans.