OCT Speckle noise reduction based on a self-supervised B2U Network

Optical coherence tomography(OCT) is a new three-dimensional tomography technology. However, the speckle noise in OCT image brings obvious limitations to its clinical application. In most real situations, it is hard to obtain high-quality OCT clean images. The self-supervised deep learning method of denoising are very popular recently, because these methods do not need clean images, and can well solve the problem that clean image cannot be obtained in real scene. In this paper, we proposed a novel self-supervised deep learning model called improved Blind2Unblind-OCT network to suppress speckle noise in OCT image. First, we improve the global-aware mask mapper based on Blind2Unblind, which can achieve better global perception in OCT images. All the sampled blind spots by mask mapper could be optimized by our designed loss function. In addition, we modify a new re-visible loss to make blind spots visible. Because all blind spots are re-visible, the OCT image will not lose important structural information. The experiments with different OCT images show that proposed model has obvious great performance compared other denoising methods of OCT image.