TM-UNet: A Deep Learning Approach for High-Resolution, Bend-Resilient Single-Fiber Imaging

Multimode fibers (MMFs) offer a promising platform for minimally invasive, high-resolution imaging due to their ability to support thousands of spatial modes. However, modal dispersion causes significant spatial distortions, resulting in speckle patterns that obscure transmitted images. Typically, these distortions can be corrected by measuring the fiber’s transmission matrix (TM) and using proximal-end spatial light modulation to perform raster-scanned imaging at the distal end. However, the TM is highly sensitive to fiber bending. To overcome this limitation, we propose a novel single-fiber imaging system that leverages pre-recorded TMs corresponding to different fiber bending conditions. During imaging, the most appropriate TM is selected based on the fiber’s shape. The acquired raster-scanned image and the corresponding TM index are processed by TM-UNet, a deep learning architecture designed to correct bending-induced artifacts and upsample 64×64 raster scans to 128×128 (two times per axis) while remaining within the fiber’s optical transfer function. TM-UNet is trained entirely on synthetic data, eliminating the need for extensive experimental calibration. Our system enables real-time image reconstruction with a fiber length of 500 mm, achieving a resolution-enhanced imaging capability at 7.04 frames per second (fps) over a range of fiber deformations up to 17.4 cm, potentially supporting live imaging applications. This approach significantly improves the robustness of MMF-based imaging, making it suitable for flexible endoscopic applications where direct distal-end access is impractical.