Unsupervised Denoising of Real Clinical Low Dose Liver CT with Perceptual Attention Networks
| Authors | Jingxi Pu et al. |
| Year | 2026 |
| Field | AI / Agents |
| arXiv | 2605.00793 |
| Download | |
| Categories | cs.AI, cs.CV |
Abstract
With the development of deep learning, medical image processing has been widely used to assist clinical research. This paper focuses on the denoising problem of low-dose computed tomography using deep learning. Although low-dose computed tomography reduces radiation exposure to patients, it also introduces more noise, which may interfere with visual interpretation by physicians and affect diagnostic results. To address this problem, inspired by Cycle-GAN for unsupervised learning, this paper proposes an end-to-end unsupervised low-dose computed tomography denoising framework. The proposed framework combines a U-Net structure for multi-scale feature extraction, an attention mechanism for feature fusion, and a residual network for feature transformation. It also introduces perceptual loss to improve the network for the characteristics of medical images. In addition, we construct a real low-dose computed tomography dataset and design a large number of comparative experiments to validate the proposed method, using both image-based evaluation metrics and medical evaluation criteria. Compared with classical methods, the main advantage of this paper is that it addresses the limitation that real clinical data cannot be directly used for supervised learning, while still achieving excellent performance. The experimental results are also professionally evaluated by imaging physicians and meet clinical needs.
Engineering Breakdown
Plain English
This paper addresses a key medical imaging problem: low-dose CT scans reduce radiation exposure but introduce significant noise that degrades diagnostic quality. The authors propose an unsupervised deep learning framework inspired by Cycle-GAN that doesn't require paired noisy/clean training data—a major practical advantage since obtaining matched CT scans is expensive and ethically complex. The framework combines U-Net for multi-scale feature extraction, attention mechanisms for intelligent feature fusion, and residual networks with perceptual loss to achieve end-to-end denoising. This approach enables hospitals to use lower radiation doses while maintaining image quality physicians need for accurate diagnosis.
Core Technical Contribution
The core novelty is adapting the unsupervised Cycle-GAN paradigm to CT denoising, eliminating the need for paired high-dose/low-dose training data that previous supervised methods required. The authors specifically combine three architectural components—U-Net's hierarchical feature extraction across multiple scales, attention mechanisms to dynamically weight important features during fusion, and residual networks to learn noise-removal transformations—into a single end-to-end pipeline. The use of perceptual loss (rather than just pixel-level L1/L2 loss) ensures denoised images look clinically realistic to radiologists, not just mathematically similar to ground truth. This unsupervised approach is substantially different from prior supervised denoising work because it trains on unpaired data, making it deployable in hospitals without expensive annotation efforts.
How It Works
The framework takes low-dose CT scans as input and processes them through a U-Net encoder-decoder architecture that extracts features at multiple scales—coarse contextual information from deep layers and fine detail from shallow layers. At each scale, an attention mechanism learns to weight which features are most relevant for denoising, suppressing noise-related features while preserving anatomical detail. Residual connections allow the network to learn the noise component directly (input minus noise equals clean image), which accelerates training since the network doesn't have to learn the identity function. The cycle-GAN component enforces cycle-consistency: a low-dose image processed through the denoiser and back through a simulated noise injection should recover the original low-dose image, creating a constraint that forces the denoiser to learn meaningful transformations rather than arbitrary distortions. Finally, perceptual loss measures how different the denoised image looks to a pretrained classifier (usually VGG), ensuring the result matches human-perceived quality rather than just pixel-level metrics.
Production Impact
For hospitals deploying this, the immediate benefit is enabling lower radiation protocols while maintaining diagnostic confidence—potentially reducing cancer risk for millions of patients undergoing CT screening. The unsupervised training approach drastically reduces implementation friction: radiologists don't need to annotate thousands of paired CT scans, and the model can be fine-tuned on institution-specific scanners without extensive relabeling. However, production deployment requires careful validation: radiologists must manually review denoised images to ensure the model doesn't hallucinate anatomy or suppress clinically important features, and you'll need A/B testing against standard reconstruction methods. Compute cost is moderate—inference on a 512×512 CT slice likely takes milliseconds on GPU, acceptable for batch processing, though real-time interactive viewing might require quantization. Integration complexity is manageable since this is a preprocessing step that sits naturally between raw CT acquisition and diagnostic workstations.
Limitations and When Not to Use This
The paper doesn't detail how well this generalizes across CT scanner vendors and protocols—a trained model from one hospital's GE scanner may fail on another hospital's Siemens equipment with different noise profiles and calibration, requiring retraining. The reliance on perceptual loss assumes that VGG features correlate with clinical diagnostic value, but this hasn't been rigorously validated against radiologist expert evaluation; the denoiser might remove subtle calcifications or small nodules that affect diagnosis. The framework assumes noise is roughly Poisson/Gaussian, but real CT noise is spatially correlated and scanner-dependent, so the unsupervised assumption of clean/noisy pairs being transformable into each other may break down at extreme dose reductions. Follow-up work needs to include prospective clinical trials showing diagnostic accuracy improvement, quantitative metrics on pathological feature preservation, and generalization testing across imaging protocols and institutions.
Research Context
This work builds directly on CycleGAN (unpaired image-to-image translation) and adapts it to the medical imaging domain where paired training data is expensive; it's in the broader trend of unsupervised medical image enhancement. Prior work like RED-NET and other supervised denoising methods required paired high-dose/low-dose scans, creating a data bottleneck that this paper sidesteps. The combination of U-Net + attention + residuals reflects the standard architectural stack now common in medical imaging, so the novelty is primarily in the training strategy (cycle consistency + perceptual loss) rather than novel layers. This opens research directions in other medical imaging modalities (MRI, ultrasound) and other dose-reduction scenarios, as well as exploring whether stronger cycle-consistency constraints or adversarial losses could further improve clinical acceptability.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::