Unsupervised Denoising of Real Clinical Low Dose Liver CT with Perceptual Attention Networks
Cycle-GAN-inspired method reduces radiation risk while preserving diagnostic quality in liver CT scans.
A team of researchers from multiple Chinese institutions has developed an unsupervised deep learning framework to denoise low-dose computed tomography (CT) scans of the liver. Published on arXiv (arXiv:2605.00793), the method addresses a critical trade-off: lowering radiation exposure for patients introduces noise that can obscure diagnostic details. Unlike most existing denoising approaches that rely on supervised learning with paired low-dose and normal-dose images, this work leverages Cycle-GAN principles to learn the mapping from noisy to clean images without paired data — a necessity for real clinical scenarios where such pairs are rarely available.
The architecture combines a U-Net for multi-scale feature extraction, an attention mechanism for effective feature fusion, and a residual network for feature transformation. Perceptual loss is added to better preserve medical image characteristics. The team also constructed a real low-dose liver CT dataset and conducted extensive experiments using both image-based metrics and evaluations by professional radiologists. Results show the method achieves excellent denoising performance while remaining clinically viable, overcoming a major limitation of prior supervised techniques.
- Unsupervised learning using Cycle-GAN avoids need for paired low-dose/normal-dose CT images, enabling direct clinical application.
- Architecture combines U-Net, attention mechanisms, residual networks, and perceptual loss for multi-scale feature extraction and fusion.
- Validation includes both standard image metrics and evaluation by imaging physicians, confirming clinical suitability.
Why It Matters
Enables safer low-dose CT screening with less radiation by denoising real clinical scans without paired training data.