Semi-supervised virtual staining using learned-illumination Fourier ptychography for high-speed label-free histopathology

Abstract Virtual staining techniques enable the digital transformation of label-free images into clinically standardized stained images. However, the high costs and time involved in generating labeled datasets for training, combined with the absence of accelerated inference pipelines for high-throughput histopathology workflows remain major challenges to their widespread adoption in clinical practice. To overcome these limitations, we present a hardware-software co-designed system that integrates high-speed Fourier ptychographic microscopy with learned illumination, supported by a semi-supervised learning framework. Our end-to-end approach employs a learned multiplexed illumination strategy that significantly reduces acquisition time while maintaining high spatial resolution across a wide field of view. On the algorithmic side, a multi-stage neural network decouples phase reconstruction from colorization, and a contrastive learning framework further generalize the virtual staining by encouraging the network to focus on intrinsic tissue features rather than absorption-induced variations. Extensive experimental results confirm the effectiveness of our method, demonstrating accurate virtual staining of label-free images while providing a scalable and cost-effective alternative to traditional histochemical staining.