PatchDenoiser: Parameter-efficient multi-scale patch learning and fusion denoiser for Low-dose CT imaging

Low-dose computed tomography (LDCT) plays a critical role in cancer screening, pediatric imaging, and longitudinal monitoring, where repeated imaging is required while minimizing radiation exposure. Despite its clinical importance, the acquisition of LDCT often introduces substantial noise and artifacts that can obscure anatomical structures, reduce diagnostic confidence, and negatively impact downstream image analysis tasks. Traditional denoising approaches based on handcrafted filtering techniques frequently over-smooth anatomical details, whereas recent deep learning methods, including CNN, GAN, and transformer-based architectures, often rely on large-scale models with high computational and energy demands, limiting their practicality for resource-constrained clinical deployment.In this work, we propose PatchDenoiser, a lightweight and energy-efficient patch-based denoising framework for LDCT restoration. The proposed architecture decomposes denoising into complementary stages: local texture extraction and global contextual aggregation, followed by a spatially aware patch fusion mechanism to preserve fine anatomical structures while effectively suppressing noise. Unlike existing large-capacity denoising networks, PatchDenoiser is specifically designed for computational efficiency, substantially reducing model complexity while maintaining competitive reconstruction quality.Extensive experiments on four LDCT datasets demonstrate that PatchDenoiser achieves comparable quantitative and perceptual performance compared to state-of-the-art CNN-based denoising methods. In particular, the proposed framework requires approximately 10 fewer parameters and 38 lower computational complexity than conventional CNN-based approaches, while maintaining comparable PSNR and SSIM performance. These results highlight the strong efficiency-performance trade-off achieved by PatchDenoiser and demonstrate its suitability for real-world clinical deployment and scalable medical imaging pipelines.Overall, PatchDenoiser provides an effective balance between reconstruction fidelity, computational efficiency, and deployment practicality, making it a promising solution for efficient LDCT denoising. The source code is publicly available at:https://github.com/JitindraFartiyal/PatchDenoiser