Mirror substrate-enhanced and physics-informed deep-learning-assisted label-free super-resolution dark-field optical...

Dark-field optical microscopy is a powerful and widely used label-free and high-contrast imaging technique. However, owing to the Abbe's diffraction limit, it is difficult to resolve fine structures at the sub-wavelength scale. Moreover, the weakly scattered signal from these fine structures can be easily submerged in background noise. In this work, we propose an optical microscopy-based label-free super-resolution imaging method that integrates mirror substrate enhancement and physics-informed deep learning. A silver-coated glass slide (mirror substrate) replaces the glass slide to hold polystyrene (PS) nanoparticles, and the scattering intensity of the nanoparticles placed on the mirror substrate is 5.9 times that on the glass slide under a dark-field optical microscopy. A physics-informed constrained convolutional neural network, which combines the initial image reconstruction strategy based on Wiener filtering with a physics-guided loss function, is developed to reconstruct the observed images. Experimental results demonstrate that, even when trained on a limited dataset, the proposed method can successfully reconstruct high-quality images of 100 nm diameter PS nanoparticle cluster samples, with a structural similarity index of 94%. This study demonstrates a dark-field optical microscope based super-resolution imaging solution that does not require fluorescent labeling or optical system modification, showing significant potential in applications such as biological imaging and nanomaterial characterization.