Single-ended Recovery of Optical fiber Transmission Matrices using Neural Networks

Ultra-thin multimode optical fiber imaging promises next-generation medical endoscopes reaching high image resolution for deep tissues. However, current technology suffers from severe optical distortion, as the fiber's calibration is sensitive to bending and temperature and thus requires in vivo re-measurement with access to a single end only. We present a neural network (NN)-based approach to reconstruct the fiber's transmission matrix (TM) based on multi-wavelength reflection-mode measurements. We train two different NN architectures via a custom loss function insensitive to global phase-degeneracy: a fully connected NN and convolutional U-Net. We reconstruct the 64 $\times$ 64 complex-valued fiber TMs through a simulated single-ended optical fiber with $\leq$ 4\% error and cross-validate on experimentally measured TMs, demonstrating both wide-field and confocal scanning image reconstruction with small error. Our TM recovery approach is 4500 times faster, is more robust to fiber perturbation during characterization, and operates with non-square TMs.