🔬 physics

Multilevel Stochastic Plug-and-Play for Sparse-View CT Reconstruction

This paper proposes Multilevel Stochastic Plug-and-Play (ML-SPnP), a novel reconstruction method for sparse-view CT that accelerates convergence and reduces runtime by performing multilevel steps in multiresolution analysis spaces to avoid the costly estimation of fine-level prior gradients, while maintaining state-of-the-art image quality.

Original authors: Antoine De Paepe, Alexandre Bousse, Dimitris Visvikis

Published 2026-06-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Antoine De Paepe, Alexandre Bousse, Dimitris Visvikis

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 solve a giant jigsaw puzzle, but you are only allowed to look at a few scattered pieces instead of the whole picture. This is what happens in Sparse-View CT (SVCT).

In a standard CT scan, the machine spins around you many times, taking hundreds of X-ray pictures from different angles to build a crystal-clear 3D image of your insides. However, this takes time and exposes you to a lot of radiation. To save time and reduce radiation, doctors sometimes only take a few pictures (sparse views).

The problem? When you try to rebuild the image from so few pieces, the result is usually a mess filled with "streaks" and noise, making it hard to see what's going on inside the body.

The Old Solution: The "Smart" Cleaner

To fix these messy images, scientists use a method called Plug-and-Play (PnP). Think of this as having a super-smart AI cleaner.

The computer makes a rough guess at the image.
It asks the AI cleaner: "Hey, does this look like a real human body? Fix the weird streaks."
The AI cleans it up, and the computer checks the math again.
They repeat this dance over and over until the image is clear.

A newer, even smarter version of this is called Stochastic PnP. Instead of just cleaning the image once, the AI adds a little bit of "static" (random noise) back in before cleaning. This helps the AI avoid getting stuck in bad guesses and makes the final image more reliable.

The Catch: This "dance" is very slow. It takes a long time to repeat the cleaning steps enough times to get a perfect picture.

The New Solution: The "Zoom-Out" Strategy (ML-SPnP)

The authors of this paper wanted to make this cleaning process much faster without losing quality. They invented a method called ML-SPnP (Multilevel Stochastic Plug-and-Play).

Here is how they did it, using a Map Analogy:

Imagine you are trying to draw a detailed map of a city.

The Old Way: You start by drawing every single tree, street sign, and house on the full-sized map immediately. You keep erasing and redrawing tiny details until it's perfect. This takes forever.
The ML-SPnP Way:
1. Zoom Out: First, you look at the map from a satellite view. You only draw the big highways and the general shape of the city. This is easy and fast.
2. Zoom In: Once the big picture is right, you zoom in a little bit. You add the main streets.
3. Zoom Further In: Finally, you zoom in all the way to add the trees and houses.

By solving the "big picture" problems first on a smaller, simpler version of the image, the computer gets a head start. It doesn't have to waste time guessing where the main structures are; it already knows them from the "zoomed-out" steps.

The Secret Trick: The Wavelet Magic

There was a tricky problem with this "Zoom-Out" idea. Usually, when you zoom out, you lose information. If you try to use the "Smart AI Cleaner" on a zoomed-out image, the AI gets confused because the "noise" it's used to seeing doesn't match the blurry, zoomed-out picture. Fixing this mismatch usually requires a lot of extra math, which slows the process down again.

The authors found a clever mathematical shortcut using something called Wavelets (think of them as a special way of organizing the image into "smooth background" and "sharp details").

They discovered that if they organized their "Zoom-Out" steps using Wavelets, the confusing mismatch between the AI and the zoomed-out image naturally disappears.

Analogy: Imagine you are trying to match two puzzle pieces. Usually, you have to sand them down and glue them together to make them fit (which takes time). But the authors found a way to cut the pieces so perfectly that they just snap together instantly without any glue.

Because of this trick, they don't need to do the expensive, slow math to fix the mismatch. They can just let the computer run the "Zoom-Out" steps very quickly.

The Results: Faster, Just as Good

The authors tested their new method on medical data (specifically, images of lymph nodes and heads).

Speed: Their method was significantly faster than the previous best methods. In some cases, it cut the time needed by more than half.
Quality: Despite being faster, the final images were just as clear and accurate as the slow methods. They didn't lose any important details.
Safety: Unlike some other AI methods that might "hallucinate" (invent fake bones or organs that aren't there), this method stays true to the actual data, which is crucial for medical safety.

Summary

The paper introduces a new way to fix blurry, low-radiation CT scans. Instead of slowly cleaning the whole image from scratch, the new method (ML-SPnP) quickly solves the "big picture" first using a special mathematical trick (Wavelets) that skips the slow, boring math steps. The result is a clear medical image generated in a fraction of the time.

Technical Summary: Multilevel Stochastic Plug-and-Play for Sparse-View CT Reconstruction

Problem Statement
Sparse-view computed tomography (SVCT) aims to reduce radiation exposure and acquisition time by limiting the number of projection views. However, this reduction renders the image reconstruction problem severely ill-posed, leading to severe streak artifacts when using traditional analytical methods like filtered backprojection (FBP). While model-based iterative reconstruction (MBIR) and deep learning-based Plug-and-Play (PnP) methods have improved reconstruction quality by incorporating learned image priors, they face significant computational challenges. Specifically, stochastic PnP methods (e.g., SNORE, Prox-SNORE) improve robustness by matching the denoiser input distribution through re-noising, but they often require many iterations to converge, limiting their practical efficiency in clinical settings. Furthermore, extending multilevel optimization (which accelerates convergence by solving problems at coarser resolutions) to stochastic PnP is non-trivial because enforcing prior coherence across resolution levels typically requires accurate estimation of fine-level prior gradients, which is computationally expensive due to the stochastic nature of the denoising steps.

Methodology
The authors propose Multilevel Stochastic Plug-and-Play (ML-SPnP), a novel algorithm that accelerates stochastic PnP reconstruction for SVCT by operating within a wavelet-based Multiresolution Analysis (MRA) framework.

Wavelet MRA Formulation: The method constructs a hierarchy of approximation spaces using 2-D Haar wavelets. The image domain is decomposed into coarse approximation coefficients and fine detail coefficients.
Stochastic Prior Coherence Cancellation: A central theoretical contribution is the observation that enforcing prior coherence across levels in a stochastic setting usually requires estimating the gradient of the stochastic prior, which involves costly Monte Carlo evaluations. The authors propose performing multilevel steps in the wavelet MRA spaces. Under an Approximation-Detail Independence Assumption (where coarse approximations and fine details are treated as independent), they prove that the stochastic prior-coherence correction term vanishes in expectation.
- Consequently, the algorithm only needs to compute the deterministic data-fidelity coherence correction, significantly reducing computational overhead.
Annealed Multilevel Updates: The algorithm employs an annealing schedule where the noise level ( $\sigma_k$ ) decreases over iterations. High noise levels in early iterations induce strong denoising for stable coarse-level corrections, while lower noise levels in later iterations recover fine details.
Algorithm Flow: The method activates multilevel updates only during an initial "warm-up" phase ( $K_{ML}$ iterations). During this phase, the algorithm traverses a hierarchy of resolutions (V-cycle), solving coarse problems approximately using stochastic fixed-point steps. After the warm-up, the algorithm proceeds with standard fine-level stochastic PnP updates (specifically Prox-SNORE), inheriting its convergence guarantees.

Key Contributions

First Stochastic Multilevel PnP: The paper introduces the first multilevel reconstruction algorithm specifically based on stochastic PnP regularization.
Theoretical Insight on Coherence: The authors identify the stochastic prior-coherence correction as the primary bottleneck for combining multilevel optimization with efficient stochastic PnP. By reformulating the problem in wavelet MRA spaces, they demonstrate that this correction vanishes in expectation, leaving only the deterministic data-fidelity correction.
Efficient SVCT Reconstruction: The proposed method achieves reconstruction quality comparable to state-of-the-art stochastic PnP methods while substantially reducing runtime.

Experimental Results
The method was evaluated on two medical CT datasets: the Lymph Nodes dataset and the Head CQ500 dataset, under sparse-view conditions (20, 30, 40, and 50 projection views).

Quantitative Performance: ML-SPnP achieved PSNR, SSIM, and LPIPS scores nearly identical to the baseline stochastic PnP methods (Prox-SNORE and Annealed Prox-SNORE).
Computational Efficiency: The method provided substantial speedups. On the Lymph Node dataset, runtime was reduced by approximately 58–67% compared to Prox-SNORE and 20–52% compared to Annealed Prox-SNORE. On the CQ500 dataset, reductions were approximately 41–57% and 6–32%, respectively.
Comparison with Diffusion and Deterministic PnP:
- Compared to diffusion-based solvers (DiffPIR, DDS), ML-SPnP offered superior perceptual quality (lower LPIPS) and fewer hallucinated structures, particularly in highly ill-posed regimes (20–30 views).
- Compared to deterministic multilevel PnP (Prox-ML-PnP), ML-SPnP sometimes showed slightly lower PSNR/SSIM but significantly better LPIPS scores and visual preservation of local anatomical details, avoiding the oversmoothing observed in the deterministic variant.
Ablation Studies:
- Prior Coherence Cost: Simulating a standard multilevel formulation without wavelet cancellation required multiple denoiser evaluations (NFE) to estimate the prior gradient, drastically increasing runtime. ML-SPnP avoided this entirely.
- Annealing: Using a decreasing noise schedule (annealing) during the multilevel phase consistently improved reconstruction quality compared to fixed-noise multilevel updates.

Significance and Claims
The paper claims that ML-SPnP offers a favorable trade-off between accuracy and computational cost for sparse-view CT reconstruction. By leveraging the structure of wavelet decompositions to eliminate the need for expensive stochastic prior gradient estimation, the method enables the use of robust stochastic PnP priors in a multilevel framework without the usual convergence penalties.

The authors position the multilevel phase as an effective warm-start mechanism rather than a standalone convergence guarantee, noting that asymptotic convergence properties are inherited from the underlying fine-level stochastic PnP algorithm. They conclude that multilevel stochastic PnP is a promising direction for efficient and reliable SVCT reconstruction, though they acknowledge limitations regarding the idealization of the independence assumption and the need for validation on real clinical data.