Physics-guided deep neural network for quantitative phase reconstruction from Zernike phase contrast microscopy

Zernike phase-contrast microscopy (ZPCM) is a widely adopted label-free imaging modality, yet it remains qualitative rather than quantitative. Unlike quantitative phase microscopy (QPM), ZPCM does not allow continuous interpretation of all spatial points on the sample and typically exhibits characteristic halo and shade-off artifacts. On the other hand, ZPCM requires simple white-light microscopy, in contrast to QPM, which necessitates more complex optical configurations and associated instrumentation costs. We present a physics-guided deep neural network for reconstructing QPM images directly from single ZPCM intensity images. The method embeds a fully differentiable, partially coherent ZPCM forward model into supervised training, explicitly modeling the formation of ZPCM intensity images from quantitative phase distributions and enforcing physical consistency. After training, the network converts a standard white-light ZPCM image into a QPM image through a single forward pass. Quantitative validation on experimental data demonstrates effective suppression of halo and shade-off artifacts and accurate recovery of phase distributions with structural similarity index (SSIM) > 0.97. Importantly, the network forward pass on a 224 × 224 image requires less than 10 msec on a standard GPU, making the approach compatible with real-time imaging workflows.