⚡ electrical engineering

Residual Diffusion with Adaptive Data Consistency for Sparse-View CT Reconstruction

This paper proposes a residual diffusion framework for sparse-view CT reconstruction that enhances sampling efficiency and image quality by initializing the reverse process from structurally informative FBP estimates, incorporating high-frequency conditioning for structural guidance, and employing an adaptive data consistency mechanism to analytically optimize relaxation parameters.

Original authors: Jia Wu, Pingting Lu, Jing Huang, Biao Qu, Zhangyong Li, Lisha Zhong

Published 2026-07-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jia Wu, Pingting Lu, Jing Huang, Biao Qu, Zhangyong Li, Lisha Zhong

Original paper licensed under CC BY 4.0 (https://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: Seeing the Invisible with Fewer Eyes

Imagine you are trying to take a photo of a complex sculpture, but you can only look at it from a few specific angles instead of walking all the way around it. If you try to piece the image together from just those few angles, the result is a blurry, streaky mess.

In medical terms, this is Sparse-View CT. Doctors usually take hundreds of X-ray images from every angle to build a 3D picture of your insides. But to save radiation (especially for kids or frequent check-ups), they want to take fewer pictures. The problem? Standard math (called Filtered Back-Projection or FBP) creates images full of "streak artifacts"—like lightning bolts of noise—that hide important details.

The Old Solution: Starting from Scratch

To fix these blurry images, scientists have started using AI Diffusion Models. Think of these models like a master restorer who has seen millions of perfect paintings.

How they usually work: The AI starts with a canvas covered in pure static (random white noise), like a TV with no signal. It then slowly "denoises" the image, step-by-step, guessing what the picture should look like based on its training.
The problem: Starting from pure noise is slow. The AI has to take hundreds of steps to turn static into a clear picture. It's like trying to sculpt a statue by starting with a pile of sand and slowly removing grains one by one. It takes forever.

The New Solution: A Smarter Starting Point

The authors of this paper propose a smarter way to start the process. Instead of starting with random noise, they start with the FBP image (the blurry, streaky one).

1. The "Residual" Shortcut (The Sketch vs. The Masterpiece)

Imagine you have a rough sketch of a face (the FBP image). It has the right shape and features, but it's messy.

The Old Way: The AI tries to turn a blank page into a masterpiece.
The New Way: The AI is trained to learn the difference (the "residual") between the rough sketch and the perfect photo. It learns to fix the specific mistakes in the sketch rather than inventing the whole image from scratch.
The Result: Because the AI starts with a good "sketch" (the FBP image) and only has to fill in the missing details, it can finish the job in far fewer steps. It's like a sculptor starting with a block of stone that is already roughly the shape of the statue, rather than starting with a pile of dust.

2. The "High-Frequency" Glasses (Keeping the Edges Sharp)

Even with the sketch as a starting point, the AI might still smooth out the fine details, like the texture of skin or the sharp edge of a bone, because it's trying to be too "safe."

The Fix: The authors added a special "glasses" module. They take the rough sketch, extract only the edges and textures (the high-frequency parts), and feed them directly into the AI's brain at every stage of the repair process.
The Analogy: It's like giving the restorer a reference photo that highlights only the sharp lines and wrinkles. This ensures the final image doesn't look like a smooth, plastic mannequin; it keeps the realistic, gritty details.

3. The "Adaptive" Compass (Balancing the Scales)

The AI needs to balance two things:

What it thinks the image should look like (based on its training).
What the actual X-ray data says (the measurements).

If the AI ignores the X-ray data, the image might look pretty but be medically wrong. If it follows the data too strictly, the streaks come back.

The Old Way: Scientists had to manually tune a "dial" (a parameter) to decide how much to trust the data vs. the AI. This is like driving a car with a manual transmission that you have to shift perfectly every time the road changes.
The New Way: The authors created an adaptive compass. At every single step of the repair, the system automatically calculates the perfect amount of trust needed based on how much error remains.
The Analogy: It's like a self-driving car that instantly adjusts its steering based on the road conditions, rather than the driver having to guess how hard to turn the wheel. This removes the need for manual tuning and prevents the AI from making mistakes when the data is very sparse.

The Results: Faster and Clearer

The paper tested this method on both computer simulations and real patient data.

Quality: The images were clearer, with fewer streaks and better preservation of fine details compared to other AI methods.
Speed: Because the AI started with a good "sketch" (the FBP image) instead of random noise, it needed far fewer steps to finish the job. This makes the reconstruction much faster, which is crucial for medical use.

Summary

The paper presents a new way to fix blurry CT scans taken with low radiation. Instead of asking an AI to imagine a picture from nothing, they ask it to fix a rough draft. They give the AI special tools to keep the edges sharp and a smart system to automatically balance the math, resulting in a high-quality image that is generated much faster than before.

Technical Summary: Residual Diffusion with Adaptive Data Consistency for Sparse-View CT Reconstruction

Problem Statement
Sparse-view Computed Tomography (CT) reduces radiation exposure and scan time by acquiring projections from a limited number of angles. However, this creates a severely ill-posed inverse problem where standard Filtered Back-Projection (FBP) reconstruction results in severe streak artifacts due to incomplete angular sampling. While deep learning methods have shown promise, supervised approaches often struggle with extremely limited views, and existing generative diffusion models face two primary constraints: (1) they typically initialize the reverse process from pure Gaussian noise, requiring a large number of sampling steps and incurring high inference costs; and (2) they often rely on fixed data consistency parameters that fail to balance the learned prior with measurement fidelity across varying sparse-view configurations.

Methodology
The authors propose a Residual Diffusion Framework tailored specifically for sparse-view CT reconstruction, comprising three core components:

Residual Forward Diffusion Process:
Instead of the standard forward process that corrupts clean images ( $x_0$ ) with Gaussian noise, the authors define a CT-specific degradation pathway. The forward process transitions from a fully sampled image to its corresponding sparse-view FBP reconstruction ( $x_{fbp}$ ). The model is trained to learn the residual $x_{res} = x_0 - x_{fbp}$ . Consequently, the reverse process is initialized from the structurally informative FBP estimate rather than random noise, significantly reducing the number of steps required to reach a high-quality solution.
Multi-Scale High-Frequency Conditioning:
To address the loss of fine details (edges and textures) inherent in FBP reconstructions, the framework introduces a conditioning module. This module extracts high-frequency features from the FBP image using a discrete wavelet transform (suppressing the approximation component and retaining detail sub-bands). These features are refined by a pre-trained auxiliary U-Net and injected into the main denoising network (a U-Net with linear attention) at multiple resolution levels. This provides explicit structural guidance to preserve anatomical boundaries and tissue textures during the reverse diffusion steps.
Adaptive Data Consistency Mechanism:
To ensure the reconstruction remains consistent with the measured projection data ( $y$ ), the authors integrate a projection-domain data consistency step at each reverse iteration. Unlike methods that use fixed relaxation parameters, this mechanism derives the relaxation parameter ( $\kappa_{t-1}$ ) analytically at each step. By solving a one-dimensional least-squares problem along the FBP correction direction using the current projection residual, the method automatically balances the generative prior against measurement fidelity without manual hyperparameter tuning. This correction is omitted at the final step ( $t=1$ ) to avoid amplifying noise.

Key Contributions

Residual Formulation: The paper reformulates the diffusion forward process as a degradation from clean images to sparse-view FBP images. This allows the reverse process to start from a physically meaningful FBP estimate, enabling significantly fewer sampling steps compared to standard Gaussian-noise diffusion.
High-Frequency Guidance: The introduction of a multi-scale conditioning module that injects wavelet-derived edge and texture features into the denoiser, specifically targeting the recovery of fine-scale anatomical structures often lost in FBP.
Adaptive Consistency: The development of an adaptive data consistency mechanism that analytically computes the optimal relaxation parameter from the projection residual at each step, eliminating the need for manual parameter selection across different sparse-view configurations.

Experimental Results
The method was evaluated on both simulated (AAPM Low Dose CT Grand Challenge) and clinical (LDCT-and-Projection) datasets.

Reconstruction Quality: The proposed method achieved the highest Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) across all tested view configurations (16, 32, and 64 views) compared to supervised baselines (FBPConvNet, LEARN++, PAIL) and other diffusion-based methods (Cold-Diff, ResShift, MCG, WSGM, VLDS).
Artifact Suppression: Qualitative results demonstrated superior suppression of streak artifacts and better preservation of fine anatomical details compared to baselines, which often exhibited over-smoothing or residual streaks.
Efficiency: The method required only 20 diffusion steps ( $T=20$ ) to achieve optimal performance, resulting in an inference time of approximately 2.8 seconds per slice. This is substantially faster than other data-consistency-guided diffusion methods (e.g., MCG, WSGM) which require hundreds or thousands of steps, though slightly slower than latent-space methods like ResShift (which the authors note suffer from over-smoothing due to lack of projection constraints).
Generalization: The method demonstrated strong cross-domain generalization, performing effectively on clinical data without retraining, attributed to the adaptive consistency mechanism adjusting to the specific projection residuals of the clinical scans.

Significance and Claims
The paper claims that by aligning the diffusion trajectory with the physical degradation process of sparse-view CT (from clean to FBP) and incorporating adaptive, projection-domain constraints, the proposed framework offers a favorable trade-off between computational efficiency and reconstruction accuracy. It addresses the specific limitations of existing diffusion models in medical imaging by reducing inference costs through structured initialization and improving structural fidelity through high-frequency conditioning and adaptive consistency. The authors note that while the method shows significant improvements, current limitations include the use of ground truth reconstructed from re-binned full-view data for clinical validation and validation only on 2D axial fan-beam geometry, with 3D cone-beam extension identified as future work.