Single-ended recovery of optical fiber transmission matrices using neural networks

Ultra-thin multimode optical fiber imaging holds promise for next-generation medical endoscopes, offering high-resolution images of deep tissues. However, current technology suffers from severe optical distortion due to fiber sensitivity to bending and temperature, requiring in vivo recalibration with single-ended access. This paper introduces a neural network (NN)-based approach to reconstruct the fiber's transmission matrix (TM) using multi-wavelength reflection-mode measurements. Two NN architectures—a fully connected NN and a convolutional U-Net—are trained using a custom loss function insensitive to global phase degeneracy. The method reconstructs 64 x 64 complex-valued fiber TMs with ≤4% error in simulations and validates results on experimental TMs. The approach enables fast, robust wide-field and confocal image reconstruction, and it is 4500 times faster and more robust to fiber perturbations than previous iterative methods. The method is also adaptable to non-square TMs.

Yijie Zheng, Terry Wright, Zhong Wen, Qing Yang, George S. D. Gordon