Abstract

Partially coherent quantitative phase imaging (QPI) underpins label-free microscopy and optical metrology, providing stable, speckle-suppressed phase measurement with a brightfield-compatible platform and resolution beyond the coherent diffraction limit. However, its image formation is governed by a high-dimensional bilinear forward model, making linearized analytic inversions such as differential phase contrast (DPC) prone to model mismatch and biased reconstruction under large phase excursions, strong absorption, and optical aberrations.

Here, we introduce USDPC, a universal neural-field solver for DPC, that performs physics-constrained, unsupervised inversion of the strict nonlinear intensity-to-phase relationship under partially coherent illumination. USDPC represents the specimen's complex transmittance and pupil aberrations as implicit neural fields and jointly optimizes them via a differentiable bilinear forward operator, enabling accurate phase recovery beyond the weak-object regime while self-calibrating unknown aberrations without additional calibration data or acquisition overhead.

Comprehensive simulations and experimental validations on microlens arrays, resolution targets, stained histological sections, and live HeLa cells demonstrate the effectiveness and universality of USDPC, and highlight physics-informed neural fields as a versatile computational framework for solving challenging nonlinear inverse problems in computational imaging and beyond.