Lensless fiber endomicroscopic phase imaging using a physical model-driven neural network

Learning-based lensless fiber endomicroscopic phase imaging through multi-core fibers (MCF) holds great promise for label-free endomicroscopic imaging of biological samples with minimum invasiveness. However, conventional data-driven deep learning approaches rely on large-scale and diverse training data, which is hard to acquire in real scenarios. To address these challenges, we propose an angular spectrum method-enhanced untrained neural network (ASNet), a training-free approach that integrates a physical model with multi-distance speckles supervision for a lensless fiber endoscope system. The feasibility of this method is demonstrated through both simulation and experiments, reflecting that ASNet can successfully resolve the USAF-1951 target with 4.38 µm resolution and achieve phase reconstruction of HeLa cells. This method enhances the robustness and adaptability of MCF-based phase imaging and serves as a versatile phase retrieval technique, paving the way for advanced applications in compact, flexible imaging systems and offering potential for clinical diagnostics.