Self-Tuning Regularization Enhances Image Scanning Microscopy

Image Scanning Microscopy (ISM) achieves super-resolution by combining detector arrays with computational reconstruction, but traditional methods like multi-image deconvolution (MID) and its sectioning variant (s^2ISM) suffer from noise amplification due to semi-convergent iterative schemes requiring early stopping.

Now, researchers have developed a self-tuning explicit regularization framework that removes the need for ad-hoc stopping. Using a Bayesian maximum a posteriori formulation with multi-frame Poisson data fidelity, they apply ℓ1 and smoothed total variation penalties. An automatic parameter selection strategy adapts the residual whiteness principle to multi-frame Poisson settings, with a spectral high-pass extension for s^2ISM. First-order optimization schemes (proximal gradient and mirror descent) with adaptive backtracking ensure stable convergence. Experiments demonstrate robust super-resolution and optical sectioning, particularly in low-photon scenarios, marking a significant step toward automated high-quality biological imaging.