High-fidelity image reconstruction in multimode fiber imaging through the MITM-Unet framework

Using a single multimode fiber for both illumination and imaging offers notable advantages in developing ultra-thin imaging probes. However, uneven speckle illumination introduces additional noise, complicating high-precision reconstruction of complex grayscale images, which remains challenging for traditional methods. In this study, we first optimize the image reconstruction framework by combining the inverse transmission matrix approach with deep neural networks, enhancing interpretability and delivering exceptional performance in reconstructing complex images. To address the noise introduced by uneven speckle illumination, we increase the target exposure and effectively integrate information from multiple illumination conditions. Results show that our proposed Multi-speckle Illumination type Inverse Transmission Matrix-Unet (MITM-Unet) method significantly outperforms the Single-speckle illumination type (SITM-Unet). Specifically, images reconstructed with MITM-Unet achieve a structural similarity index of 0.59 and a Pearson correlation coefficient of 0.91, compared to SITM-Unet's 0.38 and 0.77. These findings underscore the effectiveness of the MITM-Unet method in achieving high-quality imaging of complex grayscale targets, providing valuable insights into the imaging capabilities of single multimode fiber systems. This work holds promise for advancing simpler, more compact wide-field endomicroscopic imaging using multimode fibers.