Title: Integrating Plug-and-Play Diffusion with ADMM: Dual-Variable Coupling for Resilient Medical Image Reconstruction

Abstract:

The integration of Plug-and-Play diffusion priors (PnPDP) has established a potent framework for addressing inverse problems in imaging by leveraging pretrained generative models as modular priors. Nevertheless, significant limitations persist in current PnP solvers, such as those utilizing Half-Quadratic Splitting (HQS) or Proximal Gradient methods. These approaches operate as memoryless operators, relying exclusively on instantaneous gradients to update estimates. This absence of historical tracking results in a persistent steady-state bias, preventing reconstructions from strictly adhering to physical measurements, particularly in scenarios involving substantial corruption.

To address this deficiency, we introduce Dual-Coupled PnP Diffusion (DC-PnPDP). This method reintroduces the classical dual variable to deliver integral feedback, thereby progressively enforcing consistency between the data-consistency constraints and the prior. However, this strict geometric coupling creates a new challenge: the accumulated dual residuals manifest as spectrally colored, structured artifacts. These artifacts breach the Additive White Gaussian Noise (AWGN) assumption inherent to diffusion priors, leading to pronounced hallucinations.

To bridge this disconnect, we propose Spectral Homogenization (SH), a frequency-domain adaptation mechanism. SH transforms these structured residuals into statistically compliant pseudo-AWGN inputs, effectively aligning the solver’s rigorous optimization path with the denoiser’s valid statistical manifold. Comprehensive experiments involving CT and MRI reconstruction illustrate that our method successfully mitigates the trade-off between bias and hallucination. The results demonstrate state-of-the-art fidelity alongside significantly faster convergence. The source code is accessible at https://github.com/duchenhe/DC-PnPDP

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC