Label-free super-resolution reconstruction microscopy through illumination diversity

Super-resolution microscopy traditionally relies on fluorescent labels to achieve nanoscale imaging in the far field. Label-free modalities, including Raman scattering and coherent reflection, benefit from only a modest resolution improvement beyond the diffraction limit. While diffraction limits the information transferred by microscopes, inclusion of prior knowledge and trade-offs can be exploited to improve the spatial resolution. We show that the information obtained in image scanning microscopy (ISM) increases with off-centred and diverse illumination patterns. We use machine learning methods for data reduction and demonstrate that pixelwise reconstruction neural networks can leverage such diversity in illumination to reconstruct super-resolved images, with optional steps to further refine the reconstruction. Illumination diversity image scanning microscopy was validated experimentally with coherent reflection laser (635 and 532 nm) scanning microscopy by resolving, in real space, 300 nm and 150 nm Au nanoparticles compact islands. For the latter, Fourier analysis of the reconstructions shows a good match down to ca. 60 nm with the scanning electron microscopy ground truth. The same pixelwise method applied to incoherent fluorescence ISM of 80 nm crimson nanoparticles with excitation at 640 nm reconstructs images with a two-point separation of 190 nm, while a value of 310 nm is measured in the confocal images. The concept is noise-resistant, simple and a priori applicable to all coherent and incoherent modalities, provided useful diversity is achieved.