Abstract The pursuit of compact microscopy systems faces dual constraints from cascaded optical elements and sensor pixel limits. While the integration of metalens and sensor eliminates the bulky elements, the resolution remains confined by pixel-induced under-sampling. Here, we propose a computational imaging framework that synergizes a compact metalens microscope with a transformer-based neural network to achieve subpixel-resolution. To bridge the simulation-to-reality gap, we construct the first experimental dataset of metalens-acquired thyroid pathological sections images. The training strategy enables rapid (~ 0.2s for 110 μm × 110 μm FOV), high-fidelity (structural similarity up to 91%) reconstruction from single-frame inputs, achieving 3 × spatial sampling density with a high resolution (close to the ground truth resolution of 0.87 μm). We further demonstrate its scalability by implementing the trained network in a metalens array-based system, achieving wide-field (4 mm × 6 mm) and high-resolution (close to the Olympus 10 × /0.25NA objective) imaging, with a field of view approximately 14.5 times that of the Olympus objective. The proposed framework highlights the synergy between simplified optical hardware and computational reconstruction, paving the way for compact and intelligent microscopy.
Hu et al. (2025) studied this question.