Noise reduction in Computed Tomography is an essential factor contributing to safer clinical screening, but there is considerable noise in LDCT acquisitions that hinders visualization of anatomical structures and makes diagnostic interpretation difficult. Although many conventional denoising filters and early deep learning models are able to suppress noise, they often over-smooth the image and fail to maintain fine structural details. This study introduces Wavelet-UNet, a hybrid architecture that integrates multi-resolution wavelet decomposition with a frequency-domain cross-attention mechanism embedded within a UNet generator. The low-frequency structural content extracted via discrete wavelet transform acts as a guiding prior, enabling selective enhancement of high-frequency information and preventing the loss of clinically relevant textures. As a result, the model achieves strong noise suppression while retaining anatomical fidelity. Quantitative evaluation shows that the proposed method surpasses classical wavelet filtering and Non-Local Means, achieving a PSNR of 35.20 dB and an SSIM of 0.9635, with reduced MSE (0.00049) and MAE (0.0089), confirming its effectiveness for LDCT denoising.