ELECTRO-OPTICS  LABORATORY

Differential interference contrast microscopy (DIC) plays an irreplaceable role in live-cell dynamic studies due to its non-destructive, high-contrast, and 3D imaging capabilities. However, traditional DIC captures only single-direction gradients, causing orthogonal gradients loss and limiting quantitative phase imaging and anisotropy analysis. Here, we propose an orthogonal shear learning U-KAN (OSLU-KAN) architecture for single-direction phase gradient-based quantitative phase imaging. This method integrates highly interpretable Kolmogorov-Arnold networks (KAN) into the U-Net framework, efficiently learning to predict orthogonal phase gradients from single-direction gradients. By combining a physics-driven spiral phase integration (SPI) model and a highly compatible Fourier loss function, this method achieves fast, high-precision, and artifact-free phase reconstruction. Experimental results show an RMSE of 0.913 mrad/µm for orthogonal gradient prediction and 0.0103 rad for phase reconstruction. Importantly, OSLU-KAN enables accurate phase retrieval and anisotropic phase gradients estimation, with excellent compatibility and generalization capabilities, providing a new interpretable, physics-informed paradigm for deep learning-driven quantitative phase imaging.