Fei Xia

We introduce MicroFoundation, a universal foundation model transforming biomedical optical imaging by unifying semantic understanding with signal restoration. Addressing the fragmentation of current tools, MicroFoundation utilizes a novel architecture and was trained on over 1 million images across seven modalities and vast spatial scales (10 nm to 10 cm). It seamlessly performs diverse tasks, from segmentation to super-resolution, within a single weight set. Benchmarking demonstrates that MicroFoundation significantly outperforms existing specialized and generalist models. By offering a robust, all-in-one solution for image analysis and restoration, MicroFoundation potentially establishes a new standard for the field.

Machine learning has become a critical modern technology, but electrical hardware struggles to keep pace with the exponential growth of machine learning models. Free-space optical hardware offers a powerful alternative for large-scale computing; however, current systems are largely limited to using a single wavelength of light, restricting their processing versatility. Here, we explore the use of broadband light scattering in complex media, which generates uncorrelated optical features at each wavelength. By treating individual wavelengths as independent predictors, we demonstrate improved classification accuracy through in silico majority voting while also estimating system uncertainty without needing access to the model’s internal probability outputs. We further demonstrate that linearly combining multiwavelength features, akin to spectral shaping, enables us to tune output features with improved performance on classification tasks. These advantages of optical multiplexing pave the way toward designing more versatile hardware accelerators for next-generation artificial intelligence.

This editorial draws attention to a major issue in photonics: important research software is often hidden, undocumented, and lost over time, making results hard to reproduce. The rise of AI-generated code increases this problem by adding more ‘black box’ systems. To address this, J. Phys. Photonics is introducing Software Articles, a new article type allowing researchers to publish, validate, and share their code as formal scientific outputs. This initiative aims to promote transparency, reproducibility, and proper credit for developers. Open and peer-reviewed software helps the community verify results, reduce duplication of effort, and build lasting tools, ensuring that computational methods become reliable, accessible, and integral to scientific progress.

Dynamic imaging through time-varying scattering media is ubiquitous in real-world settings, yet it remains a defining unsolved problem as rapid spatiotemporal fluctuations overwhelm standard reconstruction pipelines that often rely on speckles with high signal-to-noise ratio. Existing approaches fall into two categories. Guidestar-based methods employ a guidestar to recover the system transfer function; however, in dynamic media, the speckle decorrelates rapidly, making the calibration quickly invalid. Guidestar-free methods infer information from speckle statistics, but rapid changes and noise often break phase retrieval. To overcome these limitations, we introduce NeOTF, a guidestar-free and neural-representation-based optical transfer function (OTF) retrieval method that enables dynamic imaging through time-varying scattering media. By optimizing this neural representation with only a few speckle images from unknown objects, NeOTF robustly retrieves the system’s OTF without a guidestar. We experimentally demonstrate robust dynamic imaging through scattering with NeOTF at an extremely low signal-to-noise ratio and broadband incoherent illumination (up to 300 nm spectral bandwidth) scenarios, and we numerically validate its dynamic imaging performance in time-varying scattering media leveraging the spatio-temporal memory effect. Finally, we discuss and analyze its computational efficiency and generalization capabilities across anisotropic scattering media. These results establish NeOTF’s promise as a practical and robust solution for dynamic imaging through scattering media.

Optical computing has attracted significant attention recently due to its potential in low energy consumption, high bandwidth, and massive parallelism. Unlike electronic systems that primarily process information sequentially with limited parallelism, light can naturally handle data in parallel both spatially and spectrally, offering strong potential for computational advancements. However, the full advantages of utilizing both the spatial and spectral properties of light in computing remain largely unexplored. In this study, we examine these benefits by using broadband light projection onto wavelength-dependent complex media. This approach approximates an additive kernel, resulting in improved performance in classification tasks compared to monochromatic light projection.

In vivo imaging of the neurovascular network is considered to be one of the most powerful approaches for understanding brain functionality. Nevertheless, simultaneously imaging the biological neural network and blood vessels in deep brain layers in a non-invasive manner remains to a major challenge due to the lack of appropriate labeling fluorescence probe pairs. Herein, we proposed a 2-photon and 3-photon fluorescence probe pair for neurovascular imaging. Specifically, the red fluorescence protein (RFP) with an absorption maximum of around 550 nm is used as a 3-photon excited probe to label neurons, and a cyanine derivative dye Q820@BSA has a NIR absorption maximum of 825 nm as a 2-photon excited probe to label the vasculature, enabling single wavelength excitation at 1650 nm for neurovascular imaging with high emission spectral separation (>250 nm). In particular, the 2-photon action cross-section of Q820@BSA was found to be about 2-fold larger than that of indocyanine green (ICG), a commonly used red 2-photon fluorescence labeling agent, at the same excitation wavelength. Benefiting from the long wavelength advantage in reducing scattering in both 2 and 3-photon excitation of the fluorescence pairs, we demonstrated in vivo neurovascular imaging in intact adult mouse brains through white matter and deep into the hippocampus in the somatosensory cortex.

Over the past decade, optical wavefront shaping has been developed to focus light through highly opaque scattering layers, opening new possibilities for biomedical applications. To probe light intensity deep inside soft scattering media such as biological tissues, internal guide-stars are required. Here, we give an overview of the main principles and describe in depth the use of a photoacoustic feedback signal for this purpose. We further present first principles calculations and simulations to estimate important experimental parameters, and detailed instructions on designing and conducting these experiments. Finally, we provide guidance towards selecting suitable equipment for building a typical experimental setup, paving the way for further innovative biomedical imaging and therapy applications.

Optical information processing and computing can potentially offer enhanced performance, scalability and energy efficiency. However, achieving nonlinearity—a critical component of computation—remains challenging in the optical domain. Here we introduce a design that leverages a multiple-scattering cavity to passively induce optical nonlinear random mapping with a continuous-wave laser at a low power. Each scattering event effectively mixes information from different areas of a spatial light modulator, resulting in a highly nonlinear mapping between the input data and output pattern. We demonstrate that our design retains vital information even when the readout dimensionality is reduced, thereby enabling optical data compression. This capability allows our optical platforms to offer efficient optical information processing solutions across applications. We demonstrate our design’s efficacy across tasks, including classification, image reconstruction, keypoint detection and object detection, all of which are achieved through optical data compression combined with a digital decoder. In particular, high performance at extreme compression ratios is observed in real-time pedestrian detection. Our findings open pathways for novel algorithms and unconventional architectural designs for optical computing.