Plug-and-Play Diffusion Meets ADMM: Dual-Variable Coupling for Robust Medical Image Reconstruction

This paper proposes Dual-Coupled PnP Diffusion with Spectral Homogenization, a novel framework that restores dual variables to eliminate steady-state bias in medical image reconstruction while transforming structured residuals into statistically compliant noise, thereby resolving the bias-hallucination trade-off and achieving state-of-the-art fidelity with accelerated convergence.

The Gist

Imagine you are trying to restore a beautiful, ancient painting that has been damaged by water, fire, and time. You have two tools to help you:

1. The Forensic Scientist (The Physics): This tool knows exactly how the water and fire damaged the painting. It can tell you, "Based on the physics of water damage, this part of the canvas must look like this." However, it's a bit rigid and sometimes gets confused when the damage is too severe, leaving the painting blurry or incomplete.
2. The Art Historian (The AI): This is a super-smart AI that has studied thousands of perfect paintings. It knows what a healthy painting should look like. It can guess missing details beautifully. But, if you let it run wild, it might start inventing things that never existed (like adding a dragon to a Renaissance portrait) because it's trying too hard to make it look "pretty."

The Problem with Current Methods
Most current methods try to use these two tools by taking a quick guess from the Art Historian, then checking it with the Forensic Scientist, and repeating.

The paper argues that current methods are like a forgetful driver. Every time they check the map (the physics), they forget where they were a second ago. They only look at the immediate error. Because they don't remember their past mistakes, they never fully correct the course. They end up driving in a circle, getting close to the destination but never quite arriving. In medical imaging, this means the final image might look okay, but it doesn't strictly match the actual data the machine collected, which is dangerous for doctors.

The Solution: The "Memory" and the "Noise Filter"

The authors propose a new system called Dual-Coupled PnP Diffusion (DC-PnP). They solve the problem with two clever tricks:

1. The "Memory Stick" (The Dual Variable)

Instead of being forgetful, they give the system a Memory Stick (called a Dual Variable).

• The Analogy: Imagine you are trying to balance a broom on your hand. If you only react to the broom's current tilt, you might overcorrect and make it fall. But if you remember how long it has been tilting (the history), you can apply the perfect amount of force to bring it back to center.
• In the Paper: This "Memory Stick" accumulates all the small errors from the past. It acts like a correction force that pushes the image until it perfectly matches the physical data. This ensures the final image isn't just "pretty"; it is mathematically guaranteed to be true to the original scan.

2. The "Noise Filter" (Spectral Homogenization)

Here is the tricky part. When the system uses its "Memory Stick" to correct the image, it creates a weird kind of static noise. It's not random static; it's patterned static (like stripes or streaks).

• The Problem: The Art Historian (the AI) was trained only on random static (white noise). If you feed it patterned static, it gets confused. It thinks the stripes are real parts of the painting and tries to "fix" them by painting over them, creating hallucinations (fake bones, fake tumors, or weird textures).
• The Solution: The authors invented a Noise Filter called Spectral Homogenization.
• The Analogy: Imagine the patterned static is a group of people shouting in a specific rhythm. The AI gets confused by the rhythm. The Noise Filter takes that rhythmic shouting and mixes in just enough random chatter to break the rhythm. Now, it sounds like a crowd of people talking randomly (white noise).
• The Result: The AI is happy! It thinks it's seeing the random noise it was trained on, so it doesn't get confused or invent fake details. It cleans up the image perfectly while the "Memory Stick" ensures the physics are still correct.

Why This Matters for Medicine

In the past, doctors had to choose between:

• Option A: An image that follows the physics perfectly but looks blurry and misses details.
• Option B: An image that looks sharp and detailed but might have fake features (hallucinations) that could lead to a misdiagnosis.

This new method gets the best of both worlds:

1. It is rigorous: It strictly obeys the laws of physics (no fake data).
2. It is sharp: It uses the AI to restore fine details without inventing fake ones.
3. It is fast: It reaches the perfect solution about 3 times faster than previous methods.

In a Nutshell:
The paper teaches a computer how to remember its mistakes (so it doesn't settle for "good enough") and how to hide its corrections (so the AI doesn't get confused and start making things up). The result is a medical image that is both scientifically accurate and visually perfect, helping doctors diagnose patients with much higher confidence.

Technical Summary

1. Problem Statement

Medical image reconstruction (e.g., CT, MRI) is an ill-posed inverse problem where high-fidelity images must be recovered from degraded, undersampled, or noisy measurements ($y = Ax + n$).

• Current State: The prevailing paradigm uses Plug-and-Play Diffusion Priors (PnPDP), which treat pretrained diffusion models as modular denoisers within iterative solvers (typically based on Half-Quadratic Splitting (HQS) or Proximal Gradient).
• The Critical Flaw: Existing PnPDP solvers function as memoryless operators (analogous to Proportional controllers in control theory). They update estimates based solely on instantaneous gradients.
  • Consequence: Under heavy corruption or ill-conditioning (e.g., limited-angle CT), these solvers fail to eliminate steady-state bias. They converge to a solution that compromises between physical data consistency and the generative prior, failing to strictly satisfy measurement constraints.
• The Proposed Solution's Challenge: Classical optimization theory suggests using Dual Variables (as in ADMM) to provide "integral action" and eliminate bias. However, reintroducing dual variables into diffusion loops creates a distributional conflict:
  • Dual variables accumulate structured, spectrally colored residuals (e.g., streaks, aliasing).
  • Pretrained diffusion models are trained strictly on Additive White Gaussian Noise (AWGN).
  • Feeding structured residuals into the diffusion model constitutes an Out-of-Distribution (OOD) input, causing the model to misinterpret artifacts as semantic content, leading to severe hallucinations.

2. Methodology: Dual-Coupled PnP Diffusion (DC-PnPDP)

The authors propose a framework that reconciles the rigorous convergence of ADMM with the statistical requirements of diffusion models through two core mechanisms:

A. Restoring the Dual Variable (Geometric Rigor)

Instead of discarding the dual variable $u^{(k)}$ (as in HQS-based methods), DC-PnPDP explicitly maintains it to track cumulative constraint violations ($x - z$).

• Role: Acts as an Integral controller in the optimization loop.
• Benefit: Theoretically guarantees asymptotic convergence to a strict consensus where the reconstructed image satisfies both the physical measurement constraints and the prior manifold, eliminating steady-state bias.

B. Spectral Homogenization (SH) (Statistical Alignment)

To prevent the structured dual residuals from triggering hallucinations in the diffusion denoiser, the authors introduce a Spectral Homogenization module.

• Mechanism: Operates in the frequency domain to transform the "colored" residuals into statistically compliant pseudo-AWGN inputs.
• Process:
  1. Diagnosis: Estimates the Power Spectral Density (PSD) of the current residual ($r = v - z$) to identify spectral deficits (where structured artifacts concentrate energy).
  2. Synthesis: Generates a complementary noise field $\xi$ that is spectrally orthogonal to the artifacts but matches the target noise level ($\sigma_t$). This fills the "spectral valleys" without destroying the "spectral peaks" (semantic content).
  3. Fusion: Adds the synthetic noise to the dual-shifted input, creating a homogenized input $\tilde{v}$ that mimics the AWGN distribution the diffusion model was trained on.
• Theoretical Guarantee: Ensures the input to the denoiser satisfies the Second-Order Spectral Consistency, making the effective noise covariance approximately a scaled identity matrix ($\sigma^2 I$).

C. Algorithm Flow

The iterative process (Algorithm 1) follows:

1. Data Consistency: Solve for $x^{(k+1)}$ using the forward operator $A$ and the dual variable $u^{(k)}$.
2. Dual Shift: Compute $v^{(k+1)} = x^{(k+1)} + u^{(k)}$.
3. Spectral Homogenization: Apply $T_{SH}$ to $v^{(k+1)}$ to generate $\tilde{v}^{(k+1)}$.
4. Denoising: Apply the pretrained diffusion denoiser $D_\theta(\tilde{v}^{(k+1)}, t)$ to get $z^{(k+1)}$.
5. Dual Update: Update $u^{(k+1)} = u^{(k)} + (x^{(k+1)} - z^{(k+1)})$.

3. Key Contributions

1. Dual-Coupled Framework: Re-introduces the explicit ADMM dual variable into PnP diffusion, proving it is essential for eliminating steady-state bias in ill-posed medical imaging tasks.
2. Spectral Homogenization (SH): Proposes a lightweight, frequency-domain adaptation mechanism that bridges the gap between the structured residuals of optimization and the AWGN assumption of diffusion priors, preventing OOD hallucinations.
3. Theoretical Insight: Provides a control-theoretic analysis showing that current PnPDP methods act as Proportional (P) controllers (prone to bias), while DC-PnPDP acts as a Proportional-Integral (PI) controller (guaranteeing convergence).
4. State-of-the-Art Performance: Demonstrates superior fidelity and convergence speed across multiple modalities.

4. Experimental Results

The method was evaluated on Sparse-View CT, Limited-Angle CT, and Accelerated MRI (using AbdomenCT-1K and fastMRI datasets).

• Quantitative Performance:
  • Limited-Angle CT (90°): Achieved +5.95 dB PSNR improvement over the strongest baseline (DiffPIR), reaching 36.98 dB vs. 31.03 dB.
  • Sparse-View CT (20 views): Outperformed the runner-up by over 3 dB.
  • Accelerated MRI (10x): Surpassed baselines by nearly 1 dB in PSNR while maintaining superior structural similarity (SSIM).
• Qualitative Improvements:
  • Successfully recovered sharp anatomical boundaries in missing wedge regions where baselines produced severe blurring and streak artifacts.
  • Eliminated coherent aliasing in MRI without hallucinating non-existent textures.
• Efficiency:
  • Achieved ~3.3x faster convergence. The method reached the performance of DiffPIR at 100 steps in just 30 steps, significantly reducing inference latency.
• Ablation Studies:
  • Confirmed that Dual Coupling alone removes bias but causes hallucinations due to spectral mismatch.
  • Confirmed that Spectral Homogenization alone offers marginal gains without the dual variable.
  • The combination of both is required to resolve the bias-hallucination trade-off.

5. Significance

This work represents a significant paradigm shift in solving inverse problems with generative models:

• Bridging Theory and Practice: It successfully integrates classical optimization wisdom (ADMM, integral action) with modern deep generative priors, solving a fundamental theoretical limitation of current PnP methods.
• Clinical Reliability: By ensuring strict adherence to physical measurement constraints (data fidelity) while maintaining high perceptual quality, the method addresses a critical need in medical imaging where "visually pleasing" but physically incorrect reconstructions are clinically dangerous.
• Generalizability: The concept of "Spectral Homogenization" offers a new pathway for adapting other structured optimization residuals to work with distribution-specific generative models, potentially applicable beyond medical imaging.