Abstract: To objectively compare denoising methods for low-dose CT (LDCT) under reproducible conditions, this study formulates and follows a unified evaluation protocol based on the public dataset LDCT-and-Projection-data. The protocol uses LD-FBP as input, full-dose (FD) images as references, unifies window width/window level and spatial resolution, adopts circular ROI as evaluation areas, and divides data by cases with consistent display standards. Four representative methods (BM3D, DnCNN, N2N-FFT, and SwinUNet) are implemented and compared. Peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and mean absolute error (MAE) are used for quantitative evaluation, supplemented by qualitative analysis including representative slice comparison, training/validation curves, and residual maps. Experimental results show that on the test set, SwinUNet and DnCNN outperform BM3D and N2N-FFT in all three indicators, with more uniform residual distribution and visual quality closer to FD images. In conclusion, supervised deep models have greater advantages in balancing noise suppression and structural fidelity, and the hybrid architecture integrated with Transformer brings a small but stable performance gain.