⚡ electrical engineering

Dynamic MRI Reconstruction Via Dual Deep Priors and Low-Rank Plus Sparse Modeling

This paper proposes a training-free dynamic MRI reconstruction framework that combines deep image priors with low-rank plus sparse modeling to effectively capture spatiotemporal correlations, achieving superior performance and theoretical convergence guarantees compared to existing supervised and unsupervised methods.

Original authors: Yongliang Sun, Siddhant Gautam, Chaoyan Huang, Nicole Seiberlich, Ismail Alkhouri, Saiprasad Ravishankar

Published 2026-05-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yongliang Sun, Siddhant Gautam, Chaoyan Huang, Nicole Seiberlich, Ismail Alkhouri, Saiprasad Ravishankar

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 Time-Lapse"

Imagine you are trying to take a high-quality video of a beating heart using an MRI machine. The problem is, MRI machines are slow. To get a clear picture, they need to collect a massive amount of data. If you try to speed up the process to get a video (which requires many frames per second), the machine has to skip a lot of data.

The result? You get a "jigsaw puzzle" with missing pieces. If you try to put the puzzle together quickly, the image looks blurry, or you see weird ghost-like artifacts (like seeing a heart in two places at once).

The Old Ways: Two Different Approaches

Before this paper, scientists tried to fix this missing puzzle in two main ways:

The "Math-Only" Approach (Compressed Sensing): This is like trying to solve the puzzle using only strict rules of logic. It's reliable, but it often leaves the picture looking a bit fuzzy or "blocky" because it doesn't know what a real heart looks like, only how the math works.
The "AI Training" Approach (Deep Learning): This is like showing a computer thousands of finished heart videos so it learns what a heart should look like. It creates amazing pictures, but it has a big flaw: it only works well if the new patient looks exactly like the patients in the training videos. If the scanner settings change or the patient is different, the AI gets confused and fails. Plus, you need a huge library of perfect, fully-sampled videos to train it, which is hard to get.

The New Solution: "The Twin Architects"

This paper proposes a new method called L+S DIP. It combines the best of both worlds without needing a massive training library.

Think of the heart video as a movie made of two layers:

The Background (Low-Rank): The parts of the heart that stay mostly the same, like the chest wall or the general shape of the heart. This is the "slow" part.
The Action (Sparse): The parts that change quickly, like the heart valves snapping shut or blood rushing through. This is the "fast" part.

The authors built a system with two specialized AI architects (neural networks) that work together:

Architect A is only allowed to draw the "Background."
Architect B is only allowed to draw the "Action."

The Magic Trick (Deep Image Prior):
Usually, AI needs to be trained on thousands of examples. But here, the architects are "untrained." They start as blank slates. However, the paper uses a clever trick: the structure of the neural network itself acts as a guide. It naturally prefers to draw smooth, natural-looking shapes (like a heart) rather than random static noise. It's like giving an artist a sketchbook that only allows them to draw organic shapes; they don't need to study a museum to know what a heart looks like.

How They Work Together: The "Extrapolated Dance"

The two architects try to fill in the missing puzzle pieces. They have to agree on the final image, but they are also being watched by a "Referee" (the data from the MRI machine). The Referee says, "Hey, your picture must match the actual data we collected, even if it's incomplete."

To make this happen fast and accurately, the paper uses a specific algorithm called eADMM.

The Analogy: Imagine two people trying to solve a maze while holding hands. If they just take small, cautious steps, it takes forever. If they take huge, reckless leaps, they might crash into walls.
The Innovation: The "Extrapolated" part of their algorithm is like a coach shouting, "You're moving in the right direction! Take a slightly bigger step in that same direction before you check the map again!" This helps them find the solution much faster without losing their way.

What They Found

The researchers tested this method on public heart MRI data.

Better than Math: It produced much clearer images than the old "Math-Only" methods, especially when the data was very incomplete (high acceleration).
Better than Standard AI: It matched the quality of the "Super-Trained AI" methods but didn't need any training data. It worked just by looking at the specific patient's scan.
The Split Works: By separating the "Background" and the "Action" into two different networks, the system could preserve the fine details of the beating heart much better than trying to draw everything with one network.

The Catch

The paper admits one downside: Because this method has to "think" and optimize for every single patient individually (like a chef cooking a meal from scratch rather than reheating a frozen dinner), it takes longer to generate the image than the pre-trained AI methods. However, for patients where no training data exists, this is a powerful tool to get high-quality images without needing a massive database of previous scans.

Summary

In short, this paper teaches a computer how to fix a blurry, incomplete heart video by splitting the job into "background" and "action" tasks. It uses two untrained AI networks that naturally prefer drawing realistic shapes, guided by a smart algorithm that speeds up the process. The result is a clear, sharp video of a beating heart, created without needing a library of thousands of past examples.

Technical Summary: Dynamic MRI Reconstruction Via Dual Deep Priors and Low-Rank Plus Sparse Modeling

Problem Statement
Dynamic Magnetic Resonance Imaging (MRI) reconstruction from undersampled k-space measurements is a challenging inverse problem. It requires preserving high spatial resolution while maintaining temporal consistency across the frames of a cine series. While classical compressed sensing methods exploit sparsity and temporal redundancy (e.g., via Low-Rank plus Sparse or L+S modeling), they rely on hand-crafted priors that may be restrictive for complex cardiac dynamics. Conversely, deep learning-based approaches achieve strong performance but typically require large, fully sampled training datasets, limiting their generalizability to new scanners, protocols, or patient populations. Training-data-free methods like Deep Image Prior (DIP) adapt to individual scans but often fail to fully exploit temporal structures and are prone to overfitting.

Methodology
The authors propose a structured framework that integrates the implicit regularization of Deep Image Prior (DIP) with the interpretability of classical Low-Rank plus Sparse (L+S) decomposition.

Dual DIP Re-parameterization: Instead of directly optimizing the image sequence, the dynamic MRI data is decomposed into a low-rank background component ( $L$ ) and a sparse dynamic component ( $S$ ). Both components are parameterized by separate, untrained 3D convolutional neural networks (CNNs), denoted as $H_{\theta_L}$ and $H_{\theta_S}$ . The Casorati matrices for these components are defined as $L(\theta_L, z_L)$ and $S(\theta_S, z_S)$ , where $z_L$ and $z_S$ are fixed random latent inputs.
Optimization Formulation: The reconstruction problem is formulated as minimizing a data fidelity term subject to low-rank and sparsity constraints on the network outputs:
$\min_{\theta_L, \theta_S} \frac{1}{2} \|Y - A(L(\theta_L, z_L) + S(\theta_S, z_S))\|_2^2 + \lambda_L \|L(\theta_L, z_L)\|_* + \lambda_S \|W S(\theta_S, z_S)\|_1$
where $Y$ is the measurement matrix, $A$ is the forward operator, $\|\cdot\|_*$ is the nuclear norm, and $W$ is a sparsifying transform.
Extrapolated ADMM (eADMM): To solve this non-convex optimization problem efficiently, the authors employ an extrapolated Alternating Direction Method of Multipliers (eADMM). The problem is reformulated with auxiliary variables $V$ (for $L$ ) and $U$ (for $S$ ). The algorithm iteratively updates these variables using singular value thresholding (for the low-rank part) and soft-thresholding (for the sparse part), followed by gradient-based updates of the DIP network parameters. Extrapolation steps are introduced to accelerate convergence.

Key Contributions

L+S Dual DIP Parametrization: The paper introduces a novel framework that explicitly models spatiotemporal dependencies by parameterizing low-rank and sparse components with distinct untrained networks. This separates slowly varying background dynamics from sparse temporal innovations without requiring supervised training data.
Convergence Analysis: The authors provide the first theoretical convergence analysis for an eADMM framework utilizing DIP-based generators. They establish a sufficient descent property using a Lyapunov function and demonstrate that every cluster point of the generated sequence is a critical point of the associated objective function, despite the non-convexity of the neural parameterizations.
Extensive Empirical Evaluation: The method is evaluated on the public OCMR dataset across various acceleration factors (4×, 6×, 8×). It is compared against classical compressed sensing, supervised learning baselines (L+S-Net), and other training-data-free methods (Multi-dynamic DIP, Low-rank DIP, and Implicit Neural Representations).

Results

Quantitative Performance: The proposed L+S DIP method consistently outperforms classical compressed sensing, other DIP variants, and Implicit Neural Representation (INR) methods across all tested acceleration factors. For instance, at 8× acceleration, it achieves a PSNR approximately 3 dB higher than the INR baseline and nearly 2 dB higher than other DIP baselines.
Comparison with Supervised Learning: The method achieves performance competitive with the supervised L+S-Net, despite not utilizing any fully sampled training data.
Ablation Studies: Experiments confirm that the dual-architecture (separate networks for $L$ and $S$ ) significantly outperforms single-network variants or models using only low-rank or only sparse priors, validating the complementary nature of the two components.
Convergence Speed: The study on extrapolation parameters ( $\alpha, \beta$ ) shows that moderate extrapolation improves convergence speed compared to non-extrapolated or overly aggressive settings.

Significance and Claims
The paper claims that this framework is particularly attractive for medical imaging scenarios where large, high-quality, fully sampled datasets are unavailable or difficult to acquire. By combining the spectral bias of DIP (which naturally favors low-frequency image structures) with explicit spatiotemporal regularization, the method effectively captures redundancy in dynamic MRI while avoiding the generalization issues of supervised deep learning. The authors emphasize that the method preserves fine anatomical structures and reduces reconstruction artifacts better than existing unsupervised approaches, all without the computational burden of training a network on external data. However, the paper modestly acknowledges that the requirement for scan-specific test-time optimization results in longer reconstruction times compared to the inference time of supervised methods.