Deep Learning for Disturbance‐Resistant Compressive Sensing Multimode Fiber Imaging

ABSTRACT Multimode fiber (MMF) is promising for minimally invasive endoscopy due to its ultra‐thin, non‐invasive nature. However, existing modulation‐based point‐scanning methods are fundamentally restricted by the intrinsic speed limits of physical modulation hardware, precluding high frame rate imaging, while compressed sensing (CS) requires a stable measurement matrix, which is difficult to maintain in MMF due to its dynamic scattering property. Here, we introduce a Physics‐integrated Disturbance‐Resistant Reconstruction Network (PDRNet). It addresses transmission matrix mismatch in CS‐MMF systems by leveraging physics priors learned from disturbed experimental data, allowing CS reconstruction at a maximum of 300 frames per second. PDRNet achieves a maximum 97.6% improvement in structural similarity (SSIM) at a 1.5% compression ratio compared to conventional CS methods on the H&E‐stained dataset. We validated these advantages on fluorescent beads, H&E‐stained tissue slices (animal/plant), and ex vivo animal tissues. By overcoming the fundamental challenge of instability, PDRNet provides MMF endoscopy with the potential for high‐speed imaging.