Physics-informed deep learning for phase retrieval and self-calibration in Zernike phase-contrast microscopy

We propose a physics-informed, self-calibrating, and non-iterative phase retrieval method for Zernike phase-contrast microscopy (ZPM) using only experimentally acquired ZPM images for training. The proposed framework jointly learns a phase-retrieval network and self-calibrates reconstruction-critical optical parameters. The training process incorporates a computationally efficient, approximation-free forward model for partially coherent ZPM image formation together with a differentiable optical parameterization. The proposed method enables unsupervised learning while ensuring physical consistency and robustness to system variations. We validate the method experimentally using COS-7 cell datasets acquired with an off-the-shelf ZPM system. This approach enables fast phase retrieval without exhaustive prior calibration or separately acquired phase data.