PatchDenoiser: Lightweight Multi-Scale Denoising for Low-Dose CT Imaging
A novel lightweight denoising system named PatchDenoiser has been introduced for low-dose CT imaging, targeting noise issues caused by lower radiation levels, patient movement, or scanner constraints. Conventional filtering techniques tend to overly smooth images, resulting in the loss of intricate anatomical features. In contrast, deep learning methods such as CNNs, GANs, and transformers often fail to maintain detail or necessitate large, resource-intensive models. PatchDenoiser separates denoising into local texture extraction and global context aggregation, combined through a spatially aware patch fusion approach. This multi-scale patch-based framework effectively reduces noise while retaining fine structural and anatomical details, presenting a practical and energy-efficient option for clinical use. The methodology is outlined in a preprint on arXiv (2602.21987).
Key facts
- PatchDenoiser is a lightweight, energy-efficient multi-scale patch-based denoising framework
- It decomposes denoising into local texture extraction and global context aggregation
- Fusion is achieved via a spatially aware patch fusion strategy
- Designed for low-dose CT images to reduce radiation exposure in cancer screening, pediatric imaging, and longitudinal monitoring
- Traditional filtering methods over-smooth and lose fine anatomical details
- Deep learning methods like CNNs, GANs, and transformers may struggle to preserve details or require large models
- The method preserves fine structural and anatomical details while suppressing noise
- Preprint available on arXiv with ID 2602.21987
Entities
Institutions
- arXiv