A physics-informed alternative to Richardson-Lucy deconvolution across SNR regimes without iteration cutoffs

Area of Science:

• Image processing
• Computational imaging
• Bayesian inference

Background:

• Richardson-Lucy deconvolution is prone to noise over-fitting and artifacts.
• Current methods require manual parameter tuning or regularization with limited physical basis.
• Robust image reconstruction is crucial in scientific imaging applications.

Purpose of the Study:

• Introduce DeBayes, a rigorous Bayesian deconvolution framework.
• Provide a principled, physics-informed alternative to existing deconvolution methods.
• Enable stable and artifact-free image reconstruction without user intervention.

Main Methods:

• Developed a Bayesian deconvolution framework (DeBayes) using a physically accurate image formation model.
• Performed deconvolution in the spatial domain, jointly modeling noise sources.
• Inferred full posterior distributions over the object, avoiding sparsity or continuity assumptions.

Main Results:

• DeBayes yields strictly positive reconstructions and converges stably without user-tuned parameters or iteration cutoffs.
• Demonstrated minimal noise amplification on simulated and experimental images.
• Achieved robust, physics-informed image reconstruction for mitochondria networks.

Conclusions:

• DeBayes offers a principled and robust alternative to Richardson-Lucy deconvolution.
• The framework provides stable convergence and minimal noise amplification.
• DeBayes is suitable for fast, parallelizable computation in scientific image reconstruction.