PSB-DSN: a two-stage unsupervised denoising enhancement network for OCT images based on pixel shuffle and blind spot masking

Optical coherence tomography (OCT) images have long been plagued by speckle noise, which has both multiplicative characteristics and spatial correlation, making noise removal extremely difficult. Traditional methods based on the Noisy2Noisy framework have poor denoising performance due to their "additive-independent" assumption. Therefore, we propose a two-stage unsupervised network, Pixel Shuffle-Blindspot Dual-Stage Network (PSB-DSN), which requires no clean data during training. In the first stage, pixel shuffle downsampling (PD) is used to cut off the noise correlation, while a global masker is introduced to select 1/4 of the pixels as blind spots. The self-replacement mechanism substitutes masked pixels with blind-spot convolution outputs, ensuring J-invariance while allowing the remaining visible pixels to provide intact contextual information. Additionally, parallel training with multiple masks is used to achieve rapid convergence and information supplementation. In the second stage, a random mask refinement network is designed to fuse the denoised output from the first stage with the original noisy image, further dispersing the checkerboard artifacts and achieving secondary enhancement of global contextual information. Evaluation on six OCT datasets shows that PSB-DSN achieves superior denoising enhancement performance over state-of-the-art unsupervised approaches. Furthermore, downstream retinal layer segmentation experiments confirm that images processed by PSB-DSN achieve consistent and stable improvements in segmentation accuracy.