Microscopy is crucial across various fields, but conventional microscopes are large, expensive, and require specialized training. This necessitates the development of compact, high-performance, and accessible microscopes. Existing miniaturized microscopes face limitations in size, performance, and cost, often compromising on resolution, depth-of-field (DOF), and chromatic aberrations. While sophisticated optical designs and multi-view information acquisition improve performance, they increase bulkiness. Recent advances in deep optics leverage parallel optimization of optical design and image processing, but they are challenged by large solution spaces and aberrations in microscopic applications. The current research addresses these limitations by developing a progressive optimization pipeline that integrates aspherical optics, computational optics, and deep learning, producing a compact, high-performance microscope that can be integrated into a cell phone.

The literature review highlights existing challenges in miniaturized microscopy. Scale-dependent geometric aberrations limit resolution, creating a trade-off between space-bandwidth product and design complexity. High resolution often reduces DOF, while chromatic aberrations further restrict applications. While miniaturization efforts have led to breakthroughs in neural recording, high-throughput screening, and flow cytometry, the optical performance of current miniaturized microscopes remains limited. Simple lens approaches result in small fields-of-view and distortions, while complex lens combinations increase size and weight. Two-photon and three-photon microscopes offer deep penetration but require specialized elements and suffer from slow acquisition speed. Deep optics technologies show promise, but current methods struggle with small working distances and large fields-of-view. The memory requirements of deep neural networks for image restoration are also substantial.

The researchers developed a progressive optimization pipeline to overcome the limitations of existing approaches. This pipeline consists of three main steps: 1) Initial optimization of an integrated lens system using traditional ray-tracing methods, focusing on field-of-view and chromatic aberrations. This step utilizes aspherical lenses and plastic optical materials to reduce chromatic aberrations and maintain compactness. 2) Integration of a diffractive optical element (DOE) featuring a cubic phase distribution to extend DOF. This step optimizes the DOE and lens system to ensure consistent modulation transfer function across a 300 µm range. 3) Concurrent optimization of the system and a deep neural network for image restoration using a simulation-supervision approach. This approach creates training data by digitally propagating defocused images through the designed microscope, generating training pairs of coded images and an all-in-focus image created using image fusion technology. A pruned deep neural network is used for real-time processing of megapixel-level captures. The entire microscope was fabricated through diamond turning and injection molding, reducing the overall cost. The PSF of the fabricated microscope was calibrated using a customized 1-µm pinhole array. Finally, the integrated microscope was integrated into a cell phone, and a pruned neural network was deployed for real-time image processing.

The progressive optimization pipeline yielded a highly compact integrated microscope (0.15 cm³ and 0.5 g), five orders of magnitude smaller than a conventional microscope. The microscope achieves 3 µm optical resolution across a 3.6 mm field-of-view and a 300 µm depth-of-field (a tenfold improvement over traditional microscopes). The simulation-supervision deep neural network for image restoration outperforms state-of-the-art shift-variant deconvolution algorithms. The integrated microscope's performance was validated using various samples, including a USAF-1951 resolution target. The device was successfully integrated into a cell phone, enabling real-time extended depth-of-field imaging. Its ability to detect skin moisture with high accuracy demonstrated its potential for portable health monitoring. The system costs less than $10 for mass production.

The results demonstrate the effectiveness of the progressive optimization paradigm in designing high-performance miniaturized imaging systems. The integrated microscope's superior performance in terms of resolution, DOF, and compactness opens up new avenues for portable diagnostics and mobile health monitoring. The simulation-supervision approach for training the deep learning model proved crucial in overcoming the challenges of ground-truth data acquisition and ensuring the network's generalization ability. The success of the cell phone integration highlights the potential of this technology for widespread applications in resource-limited settings.

This study successfully developed a highly compact and high-performance integrated microscope using a novel progressive optimization pipeline. The integration of aspherical optics, computational imaging, and deep learning resulted in a device with significantly improved depth-of-field, maintaining high resolution across a large field-of-view. The successful integration into a cell phone demonstrates its potential for various applications in portable diagnostics and mobile health monitoring. Future work could explore the use of metasurfaces to further reduce size and enhance performance, and applications beyond skin health monitoring.

While the study demonstrates impressive results, several limitations should be acknowledged. The simulation-supervision approach relies on the accuracy of the forward model, and deviations from the ideal model could affect the performance. The training data used for the neural network might not fully represent the diversity of real-world samples. The current application is limited to skin moisture detection, and further development is needed for other diagnostic applications. The generalization of the approach to other optical systems may require further investigation and adaptation.