Optimizing CT acquisition parameters to obtain diagnostic-quality images at minimal radiation dose is crucial for patient safety. However, common dose-reduction strategies, such as lowering tube voltage and current, typically degrade image quality, causing increased noise and reduced contrast. While existing deep learning methods effectively denoise low-dose CT (LDCT) images, their clinical adoption remains limited due to potential lesion loss. This limitation primarily arises because conventional pixel-, structure-, or semantic-level loss functions lack explicit constraints on realistic noise textures and high-frequency details, resulting in overly smooth outputs. To address this, we propose a Noise Power Spectrum (NPS) loss that explicitly introduces frequency-domain constraints on noise texture, effectively preserving realistic high-frequency details and mitigating over-smoothing. Additionally, we introduce a Window-Range Focused (WRF) strategy to direct the network’s attention to diagnostically relevant pixels, improving training efficiency and clinical relevance. Extensive experimental evaluations demonstrate that our approach restores LDCT images with noise and texture characteristics comparable to standard-dose CT (NDCT) images, achieving both qualitative and quantitative improvements that meet clinical standards.