Spatially variant deblur and image enhancement in a single multimode fiber imaged by deep learning

A single multimode fiber has been applied in minimally invasive endoscopy with wavefront shaping for biological research such as brain imaging. Most of the fibers, such as step-index and graded-index multimode fibers, give rise to spatially variant blur due to limits on the numerical aperture and collection efficiency. Routines to solve this problem are based on iterative algorithms, which are often slow and computer-intense. We developed a method to synthesize datasets for driving a deep learning network to deblur and denoise the spatially variant degraded image. This approach is fast (5 ms), up to three orders of magnitude faster than the iterative way. Furthermore, our method can be applied to different types of fiber endoscopy, and two types of fiber are tested here. The performance is verified on fluorescence beads and three kinds of biological tissue sections in the experiment, demonstrating effectiveness in image enhancement.