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

A research team has published a new paper, 'Plug-and-Play Diffusion Meets ADMM: Dual-Variable Coupling for Robust Medical Image Reconstruction,' proposing a novel method to fix a critical flaw in AI-powered medical imaging. Current 'Plug-and-Play' frameworks, which use pre-trained diffusion models as priors for tasks like CT and MRI reconstruction, suffer from a fundamental trade-off: they either produce images with a persistent error (steady-state bias) or generate false details (hallucinations) because the noise they feed the AI doesn't match its training assumptions. The new 'Dual-Coupled PnP Diffusion' framework directly addresses this core limitation.

The solution is a two-part innovation. First, it restores a classical 'dual variable' from optimization theory, providing integral feedback that theoretically guarantees the final reconstruction will perfectly satisfy the original physical measurements, eliminating bias. Second, it introduces a 'Spectral Homogenization' (SH) module—a frequency-domain adaptation that transforms the structured artifacts from this process into noise that statistically matches the Additive White Gaussian Noise (AWGN) the diffusion model expects. This alignment prevents hallucinations. The result, validated on CT and MRI data, is a solver that resolves the bias-hallucination trade-off, achieving state-of-the-art reconstruction fidelity with accelerated convergence, a significant step toward reliable, AI-enhanced clinical diagnostics.