Physics-guided deep learning for color Fourier ptychographic microscopy under low-frequency spectrum acquisition

Color Fourier ptychographic microscopy (FPM), particularly for high-throughput applications like digital pathology, is fundamentally limited by the trade-off between acquisition speed (sampling density) and reconstruction fidelity. To address this, we propose a novel, to the best of our knowledge, physics-guided deep learning framework for high-fidelity color FPM reconstruction under low-frequency spectrum acquisition. The core of our approach is the deep integration of the FPM physical model into a generative diffusion model, which effectively guides the model to extrapolate high-frequency spectral information beyond the captured synthetic aperture. This hybrid physical-generative approach enables robust image reconstruction from significantly reduced acquisition data, achieving high-quality results using illumination restricted to the bright-field region of the objective lens. Both simulations and experiments demonstrate that our method significantly outperforms existing traditional and deep learning-based approaches in recovering fine structural details while suppressing artifacts, without compromising computational efficiency. This physics-guided deep learning paradigm provides a novel strategy for high-speed, high-fidelity computational microscopy with significant potential for clinical translation.