Dynamic quantitative phase imaging using deep spatial-temporal prior

Non-interferometric deep learning-based quantitative phase imaging (QPI) has recently emerged as a label-free, stable, and convenient measurement of optical path length delays introduced by phase samples. Subsequently, the new paradigm of integrating deep learning techniques with physical knowledge has further enhanced the precision and interpretability without requiring a training dataset. However, this approach is often hindered by the lengthy optimization process, which severely limits its practical applications, especially for tasks that require the handling of multiple frames. In this study, we introduce a method that leverages spatial-temporal prior (STeP) from video sequences and incorporates lightweight convolutional operations into a physics-enhanced neural network (PhysenNet) for QPI of dynamic objects. Our findings indicate that we can achieve more accurate reconstructions of dynamic phase distributions without introducing additional measurements, significantly reducing both computational costs and training time by over 90%, even under low signal-to-noise ratio conditions. This advancement paves the way for more efficient and effective solutions to multi-frame inverse imaging problems.