Semi-Supervised Speckle Noise Reduction in OCT Images With UNet and Swin-Uformer
Yupei Chen, J Li, Zhongzhou Luo, Keyi Fei, Yan Luo, Zhengyu Duan, Jin Yuan, Peng Xiao
Abstract
Speckle noise is the main cause for quality degradation of Optical Coherence Tomography (OCT) images. However, speckle noise reduction is challenging due to the complex cause for statistically modelling and the requirement of a large amount of annotated data for conventional supervised learning strategies. In this paper, a novel semi-supervised learning method is proposed for speckle noise reduction in OCT images with limited labeled data. Our method creates pseudo labels for co-teaching in the training process between U-shaped Convolutional Neural Network and U-shaped Transformer with shifted window to preserve both global information and local details. The proposed scheme encourages the consistency between different streams when the advantages of both are leveraged to compensate each other for better convergence. It shows robustness on both normal and pathological OCT images with different diseases and from different devices. Our method exhibits advantages over several other state-of-the-art methods on speckle noise reduction. To our knowledge, this work is the first attempt to combine convolutional networks and Transformer for semi-supervised speckle noise reduction and achieves promising results on different datasets.