Dual blind-spot network for self-supervised denoising in OCT images

The blind-spot network and its variants have shown promising results in the field of self-supervised denoising tasks. These methods aim at concealing noisy image pixels and utilizing self-supervised learning for their restoration. However, when applied to Optical Coherence Tomography (OCT) images, which exhibits strong auto-correlation between pixels, more effective mask strategies and denoising optimization techniques are required to improve these methods. To address this challenge, this paper proposes a novel approach called Dual Blind-Spot Network (DBSN). Firstly, a fast global mask mapper is designed to break the correlation between pixels in OCT images. Inside the blind-spot network, a conditional mask convolution block with donut convolution is embedded. The OCT images are processed through the global mask mapper and pixel shuffle before being fed into the blind-spot network, achieving two different blind-spot recoveries based on corresponding conditions. Additionally, the paper discusses the lower bound of the loss function in the case of convergence and adapts changes in the weight of the loss function during training. Furthermore, a denoising refinement module is employed to improve the denoising effect during the inference stage. The effectiveness of DBSN, as a self-supervised denoising approach, is tested through numerous experiments on OCT datasets, outperforming existing methods in terms of denoising performance.