Researchers report a compact intelligent meta-microscope with subpixel-resolution and wide field-of-view

Microscopes are vital imaging tools with applications spanning scientific research, biomedical diagnosis, and industrial processes. Tao Li and colleagues from Nanjing University in China report their new intelligent meta-microscope in PhotoniX. They demonstrated a meta-microscope equipped with a neural network that achieves wide-field and subpixel-resolution imaging without hardware modification.

The miniaturization of optical microscopy holds significant promise for modern medical diagnostics, point-of-care testing, and field-portable scientific instruments. Achieving compact and light systems without compromising imaging quality is critical for broadening accessibility and enabling new applications. However, developing compact microscopy systems faces dual challenges from cascaded optical elements and sensor pixel limits. Although metalens-integrated microscopes have been proposed to eliminate bulky optical elements, the resolution remains limited by pixel-induced under-sampling. Conventional resolution enhancement methods, increasing numerical aperture (NA) or magnification, typically involve trade-offs, including larger system size, reduced signal-to-noise ratio, narrower field of view (FOV), and more complex aberration correction. Moreover, employing specialized image sensors with smaller pixel sizes requires costly hardware modifications and would reduce imaging efficiency. These issues highlight the urgent need for computational strategies that can improve the resolution of metalens-integrated microscopes without complicating their compact hardware architecture.

Researchers proposed and experimentally demonstrated a synergistic computational imaging framework that integrates a compact unit-magnification metalens microscope with a transformer-based neural network, overcoming the inherent sensor pixel size limitation through data-driven reconstruction. To bridge the gap between simulation and experimental implementation, they construct the first experimental dataset of metalens-acquired thyroid pathological sections images. The Hybrid Attention Transformer model was employed to reconstruct high-resolution images by learning the mapping from unit-magnification and three-fold magnification imaging. Once trained, the model 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 and computational subpixel-resolution (close to 0.87 μm, pixel size is 1.67 μm). A cross-tissue generalization experiment also showed that the model maintained good performance on other tissues with similar morphology.

To further extend the FOV, researchers incorporate the metalens array-based integrated microscope with the trained model, achieving wide-field (4 mm × 6 mm) and high-resolution (comparable with the Olympus 10×/0.25NA objective) imaging, providing a FOV approximately 14.5 times that of the Olympus objective with similar resolution. The thyroid pathological section imaging experiment demonstrates that the system enables clear, simultaneous visualization of normal and suspected cancerous regions within a single field of view, supporting wide-field, portable, and preliminary diagnosis of thyroid cancer for subsequent clinical evaluation.

The proposed framework highlights the synergy between simplified optical hardware and computational reconstruction, demonstrating a scalable and cost-effective approach towards compact and intelligent microscopy. By bridging advances in photonics, neural networks, and biomedical imaging, this work exemplifies the interdisciplinary integration essential for next-generation computational microscopy.