📊 statistics

Bayesian Magnetic Resonance Joint Image Reconstruction and Uncertainty Quantification using Sparsity Prior Models and Markov Chain Monte Carlo Sampling

This paper proposes a novel Bayesian framework for magnetic resonance image reconstruction that leverages sparsity priors and a proximal Markov chain Monte Carlo sampling method to simultaneously achieve superior image quality and provide reliable uncertainty quantification compared to existing optimization-based and deep learning approaches.

Original authors: Ahmed Karam Eldaly, Matteo Figini, Daniel C. Alexander

Published 2026-06-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ahmed Karam Eldaly, Matteo Figini, Daniel C. Alexander

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 are trying to solve a giant, complex jigsaw puzzle, but someone has stolen most of the pieces. You only have a few scattered pieces left (this is like the "under-sampled" data in an MRI scan). Your goal is to guess what the full picture looks like.

This paper presents a new way to solve that puzzle, not just by guessing the final picture, but also by telling you how confident you should be in that guess.

Here is the breakdown of their approach using simple analogies:

1. The Problem: The Missing Pieces

In an MRI scan, the machine collects data in a hidden format called "k-space." To get a clear picture of the brain, you usually need all the data. But collecting all the data takes a long time, which is uncomfortable for patients. So, doctors often collect only a fraction of the data (like 10% or 20%).

When you try to build the image from so few pieces, the result is often blurry or has strange ghostly lines (artifacts). Traditional computer methods try to fix this by forcing the image to look "simple" or "smooth" (using math rules called sparsity priors). They assume the brain looks like a smooth landscape or a pattern of waves, and they fill in the missing gaps based on that assumption.

2. The Old Way vs. The New Way

The Old Way (Optimization): Imagine a single expert trying to solve the puzzle. They look at the few pieces you have, apply their rules, and say, "This is the one best picture I can make." They give you a single answer. The problem? They don't tell you if they are right or if they are just guessing wildly.
The New Way (Bayesian MCMC): Instead of one expert, imagine a committee of 10,000 experts.
- Each expert looks at the few pieces and builds their own version of the puzzle based on the same rules.
- Because they are all slightly different, they come up with slightly different pictures.
- The computer takes all 10,000 pictures and averages them to get the final, most likely image.
- The Magic Part: By looking at how much the experts disagree with each other, the computer can draw a map of uncertainty. If all 10,000 experts agree on a spot, that part of the image is very reliable. If they are all arguing and drawing different things, that spot is shaky and uncertain.

3. Two Different "Rules" for the Puzzle

The authors tested two different sets of rules (mathematical models) to guide their committee of experts:

The "Smooth Edges" Rule (Total Variation): This assumes the image is made of smooth blocks with sharp edges (like a cartoon or a simple diagram).
The "Wave" Rule (Wavelets): This assumes the image is made of patterns of waves and textures (like a photograph with fine details).

They found that the "Smooth Edges" rule generally worked better for the test images they used, producing clearer pictures with fewer errors.

4. The "Confidence Map"

The most important result of this paper is the Uncertainty Map.

When the computer reconstructs the image, it also produces a second image that looks like a heat map.
Hot spots (high uncertainty): These areas correspond exactly to the places where the reconstruction is wrong or blurry.
Cool spots (low uncertainty): These are the areas where the image is clear and correct.

The authors proved that this "Confidence Map" is incredibly accurate. When they compared the map to the actual errors (by looking at the original, perfect image they were trying to reconstruct), the map predicted the errors almost perfectly.

5. Why This Beats "AI" (Deep Learning)

The paper compares their method to modern AI (Deep Learning) methods that also try to guess uncertainty.

AI Approach: Imagine training a neural network on thousands of puzzles. When it sees a new puzzle, it guesses. To guess how uncertain it is, you might train 10 different AI models and see how much they disagree. However, their disagreement might be due to the AI just being confused about its own training, not because the puzzle is hard.
This Paper's Approach: The uncertainty here comes directly from the physics of the missing data. It answers the question: "Given that we are missing 90% of the data, how much could the image possibly vary?"
The Result: The authors found their method was much better at predicting where the actual errors were compared to the AI methods. Their uncertainty maps were a much more honest reflection of the truth.

6. The "Auto-Pilot" Feature

Usually, when solving these puzzles, a human has to manually tune a "knob" (a parameter) to decide how much to smooth the image. If the knob is wrong, the picture is bad.
This new framework has a built-in auto-pilot. It automatically figures out the perfect setting for that knob while it is solving the puzzle, so no human needs to guess the settings.

Summary

The authors created a system that:

Reconstructs MRI images from very little data.
Uses a "committee of experts" (MCMC sampling) to find the best image.
Automatically adjusts its own settings.
Crucially, it produces a "confidence map" that tells you exactly which parts of the image are trustworthy and which parts are just guesses, doing a better job at this than current AI methods.

Technical Summary: Bayesian Magnetic Resonance Joint Image Reconstruction and Uncertainty Quantification

1. Problem Statement

Magnetic Resonance Imaging (MRI) reconstruction from under-sampled k-space data is an ill-posed inverse problem. While compressed sensing (CS) techniques utilizing sparsity priors (e.g., in wavelet or total variation domains) have successfully accelerated acquisition times, standard optimization-based approaches typically yield only point estimates (Maximum A Posteriori or MAP). These methods lack a principled mechanism to quantify the uncertainty associated with the reconstructed images. This absence of uncertainty quantification (UQ) limits clinical decision-making, as it is difficult to assess the reliability of specific image features or the impact of under-sampling strategies and noise. Furthermore, while deep learning methods have advanced reconstruction quality, they often introduce ambiguity regarding uncertainty due to neural network weights and may struggle with out-of-distribution data.

2. Methodology

The authors propose a novel Bayesian framework that jointly performs image reconstruction and uncertainty quantification using sparsity prior models and Markov Chain Monte Carlo (MCMC) sampling.

2.1. Hierarchical Bayesian Formulation

The reconstruction problem is modeled as a linear inverse problem $y = Ax + w$, where $y$ is the under-sampled k-space data, $x$ is the latent image, and $w$ is additive Gaussian noise.

Likelihood: Modeled as a complex Gaussian distribution.
Priors: The image $x$ $x$ is assumed to be sparse in a given transform basis $\Omega(x)$ $Ω (x)$ . The authors implement two specific priors:
1. Total Variation (TV): Promotes piece-wise smoothness, suitable for images with sharp edges.
2. Wavelet Transform: Promotes sparsity in the wavelet domain.
Hyperparameters: The regularization parameter $\tau$ is treated as an unknown random variable rather than a manually tuned constant, allowing for joint estimation within the inference process.

2.2. Inference via Split-and-Augmented Gibbs Sampling (SPAGS)

To sample from the complex, non-differentiable joint posterior distribution, the authors employ a Split-and-Augmented Gibbs Sampler (SPAGS). This approach introduces auxiliary variables to decouple the likelihood from the non-smooth prior terms (e.g., the $\ell_1$ norm of the transform).

Variable Splitting: Variables are split (e.g., $b = \Omega(x)$ , $c = Sd$, etc.) to locally decouple the optimization/inference steps.
Sampling Strategy:
- For differentiable terms (Gaussian conditionals), standard sampling is used.
- For non-smooth terms (TV or Wavelet $\ell_1$ norms), the authors utilize the Proximal Moreau-Yoshida Unadjusted Langevin Algorithm (P-MYULA). This allows efficient sampling from distributions containing non-differentiable proximal operators.
Hyperparameter Estimation: The regularization parameter $\tau$ is updated using a Stochastic Approximation Proximal Gradient (SAPG) algorithm.

2.3. Output Metrics

The framework generates a chain of samples from the posterior distribution. From these samples, the authors compute:

MMSE Estimate: The marginal posterior mean of the image pixels.
Uncertainty Map: The marginal posterior variance (or standard deviation) of the pixels, providing a pixel-wise measure of confidence.

2.4. Comparative Baselines

The proposed Bayesian methods (MCMC-TV and MCMC-Wav) are compared against:

Zero-filling IFFT: Standard inverse Fourier transform of under-sampled data.
Optimization-based MAP methods: ADMM-TV and ADMM-Wav, which maximize the log-posterior (equivalent to minimizing a cost function with variable splitting similar to the Bayesian approach but without sampling).
Deep Learning Uncertainty: Deep Ensembles (DEns) applied to MoDL and VarNet architectures, used to compare UQ quality.

3. Key Contributions

Joint Framework: Proposal of a Bayesian framework for joint MRI reconstruction and UQ using sparsity in transformed bases, a combination not previously explored with MCMC sampling in standard compressed sensing MRI.
Algorithmic Development: Development of two specific approaches (MCMC-TV and MCMC-Wav) utilizing SPAGS and P-MYULA to handle non-differentiable sparsity priors efficiently in high-dimensional image spaces.
Automated Hyperparameter Estimation: The ability to automatically estimate the crucial regularization parameter $\tau$ within the inference loop, removing the need for manual tuning or cross-validation.
Uncertainty Validation: Demonstration that the estimated uncertainty maps correlate strongly with the true error maps (computed against ground truth), validating the reliability of the UQ.
Comparative Analysis: A comprehensive study showing that the proposed Bayesian methods outperform their optimization-based counterparts (ADMM) in reconstruction quality and provide UQ capabilities that optimization methods lack. Furthermore, the Bayesian UQ shows significantly higher correlation with reconstruction error than Deep Ensemble methods.

4. Experimental Results

The framework was validated on single-coil (Shepp-Logan phantom, HCP brain images) and multi-coil (M4Raw low-field MRI) datasets using various k-space sub-sampling patterns (2D-random, Cartesian, Pseudo-radial) and ratios.

Reconstruction Quality: The MCMC-TV method consistently achieved the lowest Root Mean Square Error (RMSE) across most sampling ratios and patterns, outperforming IFFT, ADMM-TV, and ADMM-Wav. In cases of very high under-sampling (e.g., 10% or 5%), ADMM-TV occasionally showed comparable or slightly better RMSE, but MCMC-TV generally provided superior structural preservation.
Uncertainty Correlation: A critical finding is the strong correlation between the estimated uncertainty maps (marginal standard deviation) and the actual error maps.
- For the Shepp-Logan phantom, correlation coefficients ranged from 0.70 to 0.75 for MCMC-TV.
- For HCP brain images, correlations were similarly high (0.70–0.74).
- In contrast, Deep Ensemble methods (MoDL and VarNet) achieved significantly lower correlations (0.20–0.40), suggesting their uncertainty estimates are less indicative of actual reconstruction errors.
Sampling Patterns: 2D-random sampling yielded the best reconstruction quality (lowest RMSE), while Cartesian sampling resulted in the highest RMSE.
Prior Selection: TV priors generally outperformed Wavelet priors for the Shepp-Logan phantom and low-field MRI data, likely due to the piece-wise constant nature of these images. For HCP brain images, both priors performed comparably.

5. Significance and Claims

The paper claims that its primary contribution is not merely improving absolute image reconstruction quality (where deep learning often excels), but rather providing credible intervals and reliable uncertainty quantification for compressed sensing MRI.

Clinical Relevance: The strong correlation between uncertainty maps and true errors suggests that the proposed framework can serve as a reliable tool for image interpretation in clinical settings where ground truth is unavailable. It allows clinicians to identify regions of the image where the reconstruction is less certain.
Robustness vs. Deep Learning: The authors argue that Bayesian methods offer a more robust characterization of uncertainty arising from the inverse problem itself (under-sampling, noise) compared to deep learning methods, where uncertainty is often confounded by model weights and training data distribution shifts.
Automation: The framework eliminates the need for manual regularization parameter tuning, a common bottleneck in optimization-based CS methods.

The authors conclude that while computational complexity is higher than optimization-based methods, the ability to jointly reconstruct images, estimate hyperparameters, and quantify uncertainty makes the proposed Bayesian approach a powerful and principled strategy for MRI reconstruction.