⚡ electrical engineering

GroupCDL: Interpretable Denoising and Compressed Sensing MRI via Learned Group-Sparsity and Circulant Attention

This paper introduces GroupCDL, an interpretable deep learning framework that unrolls a dictionary learning algorithm with learned group-sparsity and circulant-sparse attention to achieve state-of-the-art, robust performance in both natural image denoising and Compressed Sensing MRI reconstruction.

Original authors: Nikola Janjusevic, Amirhossein Khalilian-Gourtani, Adeen Flinker, Li Feng, Yao Wang

Published 2026-02-06

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nikola Janjusevic, Amirhossein Khalilian-Gourtani, Adeen Flinker, Li Feng, Yao Wang

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

Imagine you have a very old, dusty, and scratched-up photograph. Your goal is to clean it up so you can see the details clearly again. This is what "image denoising" is: taking a messy picture and making it clean.

For a long time, computers have been good at this, but they work like "black boxes." You put a dirty photo in, and a magic box spits out a clean one. You get a great result, but nobody inside the box knows why it worked or how it decided what to keep and what to throw away. It's like a chef who makes a perfect stew but won't tell you the recipe.

This paper introduces a new method called GroupCDL. Think of it as a chef who not only makes a perfect stew but also writes down the exact recipe, explaining every step. Here is how they did it, using some simple analogies:

1. The "Group" Idea: Finding Look-Alikes

Imagine you are trying to fix a torn map. If you look at a patch of the ocean on the left side of the map, you might find a patch on the right side that looks exactly the same (both are blue water with similar waves).

Old Way: Most computer programs look at one tiny square of the image at a time. They try to fix that square without really caring about the rest of the picture. It's like trying to fix a single puzzle piece without looking at the picture on the box.
The GroupCDL Way: This new method is smart enough to say, "Hey, this patch of sky looks just like that other patch of sky over there!" It groups these similar-looking patches together. If one patch is noisy (blurry), it looks at its "twin" to figure out what it should look like. This is called Nonlocal Self-Similarity.

2. The "Group Threshold" Filter: The Smart Sieve

Once the computer finds these groups of similar patches, it needs to decide what to keep and what to throw away.

The Old Filter: Imagine a sieve that just shakes everything. If a grain is too small, it falls through; if it's big, it stays. This is a simple "on/off" switch.
The New Filter (Learned Group-Thresholding): The new method uses a "smart sieve." It looks at the whole group of similar patches at once. If most of the group agrees that a certain detail is important, the sieve keeps it. If the group agrees it's just noise, the sieve throws it away. It's like a committee of experts voting on what belongs in the picture. This makes the cleaning process much more accurate.

3. The "Circulant-Sparse Attention": The Efficient Librarian

Here is the tricky part. If you have a huge photo (like a high-definition MRI scan), finding every single matching patch in the whole image would take forever. It's like trying to find a specific book in a library by checking every single book on every single shelf one by one.

The Problem: Traditional methods try to check everything, which is slow and uses too much memory.
The Solution (Circulant-Sparse Attention): The authors invented a clever shortcut. Imagine the library is arranged in a circle. Instead of checking every book, the librarian only checks the books in the immediate neighborhood (a small window) but does it in a way that wraps around the edges of the room.
- They call this Circulant-Sparse Attention.
- It's like looking at a neighborhood block rather than the whole city, but doing it so efficiently that it feels like you saw the whole city. It allows the computer to handle huge images (like medical scans) without crashing or taking hours to process.

4. Why This Matters: The "Transparent" Machine

The biggest win of this paper isn't just that the pictures look good (they do, as good as the best "black box" methods). The win is transparency.

Because the authors built this system by mathematically "unrolling" a known cleaning algorithm, they know exactly what every part of the computer is doing.

No Black Boxes: We know why it removed the noise.
Robustness: If the computer is trained on "medium" noise but has to clean a "very noisy" picture later, it doesn't panic. It adapts well because it understands the underlying math, not just memorized patterns.

5. The Medical Application: MRI Scans

The paper also shows this works for MRI scans (Compressed Sensing MRI).

The Analogy: Imagine trying to take a photo of a moving car, but your camera is broken and can only take 1/4 of the picture. You have to guess the rest.
The Result: GroupCDL is excellent at guessing the missing parts of the MRI scan because it uses those "group" rules to fill in the gaps accurately. It produces clearer medical images than current top-tier methods, and because it's transparent, doctors can trust the process more.

Summary

In short, the authors built a new image cleaner that:

Groups similar parts of an image together to help each other.
Uses a smart voting system (group-thresholding) to remove noise.
Uses a super-fast, circular search method (CircAtt) to handle huge images without slowing down.
Is fully explainable, meaning we know exactly how it works, unlike the "magic black boxes" used today.

It's like upgrading from a magic trick to a masterclass in image cleaning.

Technical Summary: GroupCDL: Interpretable Denoising and Compressed Sensing MRI via Learned Group-Sparsity and Circulant Attention

Problem Statement
Nonlocal self-similarity (NLSS) is a powerful prior for image restoration, yet state-of-the-art deep learning methods utilizing NLSS are typically constructed as "black-box" architectures. These models often lack interpretability, making analysis and improvement difficult. Furthermore, current implementations frequently rely on processing overlapping image windows independently (Patch-based Dense Attention, or PbDA), which neglects dependencies between overlapping regions and incurs computational redundancy. While previous interpretable networks (e.g., CDLNet) achieved robustness to noise-level mismatches, they utilized local $\ell_1$ sparsity priors and failed to bridge the performance gap with black-box nonlocal models. Additionally, existing nonlocal interpretable approaches have struggled to achieve competitive performance or handle global artifacts effectively in large-scale inverse problems like Compressed Sensing MRI (CS-MRI).

Methodology
The authors propose GroupCDL, an interpretable convolutional neural network derived from unrolling a convolutional dictionary learning algorithm with a group-sparsity prior. The architecture addresses the limitations of previous methods through two primary innovations:

Learned Group-Thresholding (GT):
Instead of the standard element-wise soft-thresholding used in $\ell_1$ -based denoising (like CDLNet), GroupCDL employs a learned group-thresholding operator. This operator encodes the NLSS prior by performing spatially varying soft-thresholding on the latent representation. The thresholding is dictated by an adjacency matrix $\Gamma$ that measures the closeness between latent pixels. Unlike classical group-sparsity where $\Gamma$ is pre-computed and binary, GroupCDL learns a "soft" similarity matrix $\Gamma$ end-to-end. This is achieved via learned pixel-wise transforms ( $W_\theta, W_\phi$ ) that compute similarities within local windows, followed by a row-wise softmax normalization.
Circulant-Sparse Attention (CircAtt):
To enable efficient training and inference on large images while maintaining the translation-equivariance of convolutional sparse coding, the authors introduce CircAtt.
- Mechanism: CircAtt computes similarities only within a sliding window of size $W \times W$ with circular boundary conditions. This results in an adjacency matrix with a Block-Circulant with Circulant Blocks (BCCB) sparsity pattern.
- Efficiency: By restricting computations to local windows and utilizing the BCCB structure, the method avoids the quadratic memory and computational complexity ( $O(Q^2M)$ ) of dense attention. Instead, it achieves linear complexity relative to the number of pixels ( $O(QW^2M)$ ), where $Q$ is the number of latent pixels and $M$ is the channel count.
- Global Artifacts: Unlike patch-based methods that process windows independently and average overlaps (which can cause boundary artifacts), CircAtt processes the entire image jointly in a single forward pass, allowing for the effective restoration of global degradations such as aliasing artifacts in CS-MRI.

The network is formulated as an unrolled Proximal Gradient Method (PGM) for the Basis Pursuit DeNoising (BPDN) problem. For CS-MRI, the architecture is extended by incorporating the observation operator $H = MFR$ (masked Fourier transform with coil sensitivities) and its adjoint, along with a Gram operator $H^H H$ , within each unrolled layer.

Key Contributions

Novel Nonlocal Operation: The introduction of Circulant-Sparse Attention (CircAtt), which addresses the modeling and computational shortcomings of common Patch-based Dense Attention (PbDA) by maintaining translation equivariance and enabling joint processing of large images.
Learned Group-Thresholding: A differentiable group-thresholding operation that encodes the group-sparsity prior using a reduced channel dimension for the latent space, achieving competitive inference speeds.
Interpretable Performance: The construction of an interpretable nonlocal CNN that achieves denoising and CS-MRI reconstruction performance on par with state-of-the-art black-box models (including Transformers and nonlocal DNNs) without relying on common deep-learning tricks like batch normalization or residual learning.
Robustness: The method demonstrates novel robustness to noise-level mismatches between training and inference, a property inherited from the underlying BPDN formulation.
Open Source: The authors provide an open-source implementation in the Julia programming language.

Results
Experimental results indicate that GroupCDL achieves competitive natural-image denoising performance compared to black-box nonlocal DNNs and transformers. In the domain of CS-MRI, the method yields state-of-the-art performance, outperforming existing methods. The study highlights that the interpretable construction allows for straightforward extension to inverse problems without the "tricks" often required to adapt black-box methods. Furthermore, the use of CircAtt enables effective restoration of global artifacts (e.g., aliasing) that are difficult to handle with patch-based approaches.

Significance
The paper claims that GroupCDL represents the first instance of an interpretable network that successfully bridges the performance gap to state-of-the-art nonlocal black-box methods for image denoising. By deriving the network directly from a group-sparse convolutional dictionary learning formulation, the authors demonstrate that high performance does not require sacrificing interpretability. The proposed architecture offers a principled alternative to black-box designs, providing parameter and dataset efficiency, novel generalization capabilities (specifically regarding noise-level mismatches), and a unified framework for both image denoising and CS-MRI reconstruction. The work challenges the necessity of black-box constructions for achieving nonlocal self-similarity benefits, showing that structured, interpretable designs can match or exceed them while offering superior robustness and efficiency in handling global image degradations.