Utilizing U-Nets for quantitative differential phase contrast imaging for biomedical images

Abstract

Traditional quantitative differential phase-contrast (qDPC) imaging is constrained by the need for multiple intensity measurements, which limits its application for real-time living cell monitoring. This study introduces a single-shot quantitative DPC imaging method that integrates radially asymmetric color-encoded illumination with a specialized deep learning framework. The core architecture consists of two sequential U-Net models designed to perform a sophisticated intensity-to-phase translation. The first U-Net is trained to generate an initial phase estimate from single-shot RGB intensities. The second U-Net then refines this output into a high-quality, isotropic quantitative phase image. This two-stage approach was found to be superior to end-to-end networks because it specifically mitigates phase errors caused by sample dispersion and chromatic aberration. The framework was developed using a dataset of thirteen cell lines, with 264 images used for training over 500 epochs using the Adam optimizer and mean squared error (MSE) loss. The model utilizes high-accuracy phase retrieved from linear-gradient pupils as the ground-truth target. Experimental results demonstrate that the DL-based method achieves a Structural Similarity Index (SSIM) of >0.98 and keeps phase differences within 13% of the ground truth. With an inference time of approximately 0.5 seconds for large-scale images, this deep learning approach effectively bypasses the hardware latency of traditional multi-shot systems, enabling high-speed quantitative phase retrieval.