Physics-embedded neural state-space module for single-shot lens-free on-chip microscopy

Lens-free on-chip microscopy (LFOCM) enables large-field-of-view, high-throughput quantitative imaging by capturing diffraction holograms directly at the sensor plane, eliminating the need for bulky optics and mechanical scanning. Conventional reconstruction pipelines, however, suffer from inherent noise amplification due to their primary reliance on physics-based priors. To overcome this limitation, we developed a physics-embedded neural state-space framework that incorporates diffraction physics for single-shot LFOCM reconstruction. This physics-constrained approach integrates a dual-branch architecture that combines local inductive bias with global modeling, enabling unsupervised recovery of high-fidelity quantitative amplitude and phase distributions. Comprehensive validation demonstrates state-of-the-art noise robustness and reconstruction fidelity, achieving simultaneous artifact suppression and resolution preservation.