Quantitative phase imaging (QPI) is invaluable for non-invasive, label-free phase retrieval from biological samples. However, traditional phase reconstruction from a single off-axis digital hologram (ODH) requires a two-step process consisting of autofocusing followed by phase reconstruction. This conventional approach has significant limitations, including a restricted initial defocus range, an inefficient iterative focal plane search, and susceptibility to noise, all of which hinder robust and accurate phase recovery. This study introduces a novel and efficient conditional generative adversarial network (CGAN)-based method to address these challenges by integrating autofocusing with phase reconstruction from ODHs.

Our network architecture incorporates skip connections for multi-scale feature fusion, preserving fine image details and enhancing adaptability to a broader range of defocusing scenarios. To validate our method, an off-axis digital holographic microscopy interferometry system was constructed, and a dataset was compiled using PC3 cell samples for training and testing. This dataset, representing a common biological cell line, provided a realistic benchmark across various defocus levels and sample morphologies. The network learns a direct, end-to-end mapping from a defocused hologram to its corresponding focused phase map, bypassing the complex and time-consuming iterative optimization processes of traditional methods.

Experimental results demonstrate the superior performance of the proposed CGAN-based method compared to established architectures such as ResNet-pix2pix, U-Net, and CycleGAN. Quantitative metrics, including the structural similarity index measure, peak signal-to-noise ratio, and root mean square error of phase profiles, consistently favored our approach, indicating a significant improvement in accuracy and robustness. The multi-scale feature fusion mechanism allows the network to capture both low-level textural details and high-level structural information, improving its ability to handle varying degrees of defocus and noise, while eliminating the need for iterative focal plane searching and explicit aberration compensation.

The proposed CGAN-based framework offers a streamlined and powerful solution for QPI of biological specimens by integrating autofocusing and phase reconstruction into a single, trainable network. This represents a significant advancement, providing a highly efficient and robust phase reconstruction framework for digital holographic phase imaging. The ability to accurately recover phase information without iterative procedures and manual parameter tuning makes the framework ideal for applications requiring rapid acquisition and analysis, such as live-cell imaging and high-content screening. Its inherent robustness to noise further extends its applicability to suboptimal imaging conditions. This research lays the groundwork for developing more sophisticated and automated phase imaging systems. It paves the way for new discoveries in cell biology, drug discovery, and diagnostics, and represents a significant step forward in realizing the full potential of digital holography for label-free QPI.