“…13 Current deep learning models, however, focus on restoring simple geometric patterns and "one-to-one" mapping datasets. To overcome MMF's variability and randomness in continuous transmission and imaging, scalable semi-supervised learning model, deep de-speckle complex neural network, convolutional neural network (CNN), and multi-scale memory dynamic-learning network have been designed by Fan, 14 Liu, 15 Xu, 16 and Li et al 17 While ARTICLE pubs.aip.org/aip/adv these networks have achieved transient and short-term continuous transmission and imaging for complex images through MMFs, a more robust and long-term MMF de-speckle neural network is still needed for practical scenarios, such as long-term visualization and minimally invasive surgery. This paper presents a multi-channel symmetric network (MCSNet) architecture, comprising U-Net 18 and ConvNeXt Block, 19 to predict the effective features of speckles.…”