⚡ electrical engineering

Unsupervised Denoising of Real Clinical Low Dose Liver CT with Perceptual Attention Networks

This paper proposes an unsupervised deep learning framework combining U-Net, attention mechanisms, and perceptual loss to effectively denoise real clinical low-dose liver CT scans, overcoming the limitations of supervised learning on real data while achieving performance validated by medical experts.

Original authors: Jingxi Pu, Tonghua Liu, Zhilin Guan, Siqiao Li, Yang Ming, Zheng Cong, Wei Zhang, Fangwei Li

Published 2026-05-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jingxi Pu, Tonghua Liu, Zhilin Guan, Siqiao Li, Yang Ming, Zheng Cong, Wei Zhang, Fangwei Li

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: The "Blurry" X-Ray

Imagine you are trying to read a map in the dark. To see clearly, you need a bright flashlight. But in the world of medical CT scans, a bright flashlight (high radiation) is dangerous for the patient, especially for children or women who are more sensitive to it.

So, doctors use a dimmer flashlight (Low-Dose CT, or LDCT). The problem? The image comes out grainy and noisy, like a photo taken in the dark with a shaky hand. Important details, like a small tumor or a lesion, get lost in the "snow" of the image. If a doctor can't see the details clearly, they might miss a diagnosis.

The Old Way vs. The New Way

The Old Way (Supervised Learning):
Usually, to teach a computer how to clean up a blurry photo, you show it a "before" picture (blurry) and an "after" picture (clear). The computer learns the difference.

The Catch: In the real world, you can't take two pictures of the same patient at the exact same time—one with a dim flashlight and one with a bright one. The patient would move, or they would get too much radiation. So, for real patients, we don't have these perfect "before and after" pairs. The old methods get stuck because they need this perfect data.

The New Way (This Paper's Solution):
The authors built a smart system that learns without needing perfect pairs. Think of it like a translator who learns two languages just by listening to people speak, without a dictionary.

How the System Works (The "Magic Translator")

The researchers created a framework based on Cycle-GAN (a type of AI that learns to translate styles). Here is how their specific "translator" is built:

The U-Net (The Multi-Layered Sifter):
Imagine a sieve that catches different sizes of sand. The system uses a structure called U-Net that looks at the image at different "zoom levels." It catches the big shapes (like the outline of the liver) and the tiny details (like a small blood vessel) all at once. This ensures nothing important is lost.
The Attention Mechanism (The Spotlight):
When the system mixes information, it uses an Attention Module. Think of this as a spotlight in a dark room. Instead of looking at everything equally, the spotlight tells the AI, "Hey, pay attention to this specific part of the image, and ignore the rest." It helps the AI focus on the important medical details and ignore the noise.
The Residual Network (The Refiner):
The system uses Residual Blocks to refine the image. Imagine a sculptor who doesn't start from a block of stone every time; instead, they take a rough shape and chip away small bits to perfect it. This helps the AI transform the noisy image into a clean one step-by-step without losing the original structure.
Perceptual Loss (The "Human Eye" Judge):
Standard math checks if pixels match exactly. But the human eye doesn't work that way; we care about texture and feeling. The authors added Perceptual Loss, which uses a pre-trained AI (VGG-19) to act as a "critic." It doesn't just check if the pixels are right; it checks if the image feels like a real, high-quality medical scan.

The "2.5D" Trick (Using Memory Wisely)

CT scans are actually stacks of 2D slices (like a loaf of bread).

The Challenge: Looking at just one slice ignores the context of the slices above and below it. But looking at the whole 3D loaf at once requires too much computer memory.
The Solution: The authors used a 2.5D approach. Imagine looking at three slices at once (the one you are fixing, plus the one above and below it) to get context, but processing them one by one.
Transfer Learning: They trained the AI on 2D slices first, then "migrated" that knowledge to the 3D task. It's like learning to drive a car on a simulator (2D) before driving the real car (3D). This saved them about 80% of the training time.

The Results: What Did They Prove?

The team tested their system in two ways:

Standard Test (Mayo Dataset): They used a public dataset where they did have perfect pairs. Here, their method performed just as well as the best existing methods, proving the math works.
Real World Test (Mudanjiang Hospital): This is the big one. They used real liver scans from patients where they didn't have perfect pairs.
- The Outcome: The system successfully cleaned up the noisy, low-dose images.
- The Proof: They showed the results to experienced radiologists. The doctors said the cleaned-up images looked clear enough to spot lesions (abnormalities) that were previously hidden in the noise.

The Bottom Line

This paper introduces a way to clean up low-radiation liver CT scans without needing perfect training data. By combining a multi-scale filter (U-Net), a spotlight for focus (Attention), and a "human-eye" judge (Perceptual Loss), they created a tool that helps doctors see clearly even when the radiation dose is kept low to protect the patient.

They also built a new, real-world dataset of these scans to help other researchers in the future.

1. Problem Statement

Context: Computed Tomography (CT) is vital for clinical diagnosis but exposes patients to ionizing radiation. To adhere to the ALARA (As Low As Reasonably Achievable) principle, Low-Dose CT (LDCT) is increasingly used. However, reducing X-ray intensity introduces significant noise and artifacts, which can obscure lesions, hinder radiologists' judgment, and lead to misdiagnosis.

The Core Challenge:

Data Scarcity for Supervised Learning: Traditional deep learning denoising requires paired datasets (LDCT images and their corresponding ground-truth Normal-Dose CT (NDCT) images of the same patient). Acquiring such pairs is clinically impossible because scanning the same patient twice at different doses doubles the radiation exposure.
Real-World Limitations: Existing unsupervised methods often fail to effectively remove noise from real clinical data or struggle with the complex mixture of Poisson, Gaussian, and multiplicative noise found in actual LDCT scans.
Visual Fidelity: Many existing methods (e.g., RED-CNN) tend to over-smooth images, removing noise but also erasing critical fine details and textures necessary for diagnosis.

2. Methodology

The authors propose an end-to-end unsupervised LDCT denoising framework based on a modified Cycle-Generative Adversarial Network (Cycle-GAN) architecture, specifically tailored for medical imaging.

A. Network Architecture

The framework consists of two generators ( $G$ and $F$ ) and two discriminators ( $D_X$ and $D_Y$ ) operating in a cyclic loop:

Generator ( $G$ ): Transforms LDCT to NDCT.
- Structure: A hybrid architecture combining U-Net, Residual Networks (ResNet), and Attention Mechanisms.
- Encoder-Decoder: Uses a U-Net structure (downsampled 8 times) to extract multi-scale features, preserving both shallow details and deep semantic information.
- Residual Blocks: Nine ResNet blocks are placed in the bottleneck to facilitate feature transformation and prevent vanishing gradients.
- Attention Module: A custom attention mechanism is inserted at the feature fusion stage (skip connections). It calculates weights to effectively fuse LDCT and NDCT features, suppressing redundant information and enhancing relevant features.
Discriminators: Utilize a PatchGAN (70×70) structure to perform fine-grained, local discrimination, ensuring high-frequency texture fidelity rather than just global consistency.

B. Loss Functions

The total loss function ( $L_{total}$ ) combines three components to balance realism, consistency, and perceptual quality:

Adversarial Loss ( $L_{adv}$ ): Drives the generated images to match the real data distribution (NDCT style).
Cycle-Consistency Loss ( $L_{cyc}$ ): Ensures that $G(F(x)) \approx x$ and $F(G(y)) \approx y$ , preserving content structure in the absence of paired data.
Perceptual Loss ( $L_{perc}$ ): A critical innovation. Instead of pixel-wise comparison, it uses a pre-trained VGG-19 network to extract features from the lower 16 layers. This loss minimizes the difference in high-level feature space, ensuring the denoised image retains the texture and visual characteristics of a real NDCT scan.

C. Transfer Learning & 3D Processing

2.5D Design: To leverage 3D spatial context without the high computational cost of full 3D convolutions, the model processes three adjacent slices simultaneously (2.5D).
Parameter Migration: The authors employ transfer learning. A model trained on 2D slices is used to initialize a 3D-aware model. The middle layer parameters of the 3D model are initialized with the trained 2D weights, while other layers are set to zero. This reduces training time by ~80% and allows the model to learn from 3D contextual information (e.g., vessel continuity) while maintaining in-plane resolution.

3. Key Contributions

Unsupervised Framework for Real Data: The paper addresses the critical limitation of real clinical data by proposing an unsupervised method that does not require paired LDCT/NDCT datasets, making it applicable to real-world hospital scenarios.
Hybrid Architecture: The integration of U-Net (multi-scale extraction), ResNet (feature transformation), and a custom Attention mechanism (weighted feature fusion) creates a robust generator specifically for liver CT denoising.
Perceptual Loss Integration: By incorporating VGG-based perceptual loss, the method overcomes the "over-smoothing" issue common in pixel-based loss functions, preserving diagnostic textures.
Real Clinical Dataset: The authors constructed and utilized a real-world liver CT dataset from Mudanjiang Second Affiliated Hospital, containing various low-dose protocols (80kV, 100kV) and slice thicknesses (1mm, 5mm), distinct from synthetic datasets like AAPM-Mayo.
Efficient 3D Learning: The proposed 2.5D transfer learning strategy effectively balances computational cost with the need for 3D anatomical context.

4. Experimental Results

The method was evaluated on two datasets: the standard AAPM-Mayo dataset (paired) and the Real Clinical Liver Dataset (unpaired).

Quantitative Metrics (Mayo Dataset):
- The proposed method achieved a PSNR of 29.176 and SSIM of 0.9076, outperforming standard CycleGAN variants and competing closely with supervised baselines like WGAN-VGG.
- Crucially, it achieved a significantly lower Perceptual Loss (2.63) compared to RED-CNN (4.26) and WGAN (4.02), indicating superior texture preservation.
Qualitative & Clinical Evaluation:
- Visual Quality: The method successfully removed noise while keeping lesions and fine structures clearly visible (unlike RED-CNN which blurred details).
- Radiologist Assessment: Two senior radiologists performed blind subjective evaluations on a 5-point scale. The denoised images were judged to meet clinical diagnostic needs, with lesions in the Region of Interest (ROI) clearly distinguishable from noise.
- Robustness: The model demonstrated universality, successfully denoising 100kV and 5mm slice thickness data, though performance varied slightly with lower resolution inputs.
Radiation Dose Analysis: The study confirmed that 80kV scanning significantly reduces radiation exposure (CTDIvol, DLP, and Effective Dose) compared to 120kV, validating the necessity of the denoising technology to make low-dose protocols viable.

5. Significance

This paper represents a significant step forward in the practical application of AI in medical imaging.

Clinical Viability: By solving the "paired data" bottleneck, it enables the deployment of deep learning denoising in real hospitals where paired datasets do not exist.
Diagnostic Safety: The preservation of fine details and textures via perceptual loss ensures that the denoising process does not compromise diagnostic accuracy, addressing a major safety concern in AI-assisted radiology.
Efficiency: The transfer learning approach makes the training of complex 3D-aware models feasible on standard clinical hardware, lowering the barrier to entry for medical AI research.

In summary, the authors have developed a robust, unsupervised solution that effectively transforms noisy low-dose liver CT scans into high-quality images suitable for clinical diagnosis, bridging the gap between theoretical deep learning models and practical medical application.