The Research Of OCT Image Restoration Based On Deep Learning

Optical Coherence Tomography(OCT)is an imaging modality that uses optical interference to perform tomographic scanning of samples.It has been widely used in medical diagnosis and auxiliary therapy due to its characteristics of harmless,large imaging depth,high-resolution,strong tomographic ability,and so on.However,there are two main issues that hinder the further development of OCT technique:1)the speckle noise introduced by low-coherence interference.OCT imaging system adopts a broadband light source and the speckle noise will inevitably be introduced due to the phase difference when imaging is performed using the optical interference,resulting in low quality of the captured OCT image,which affects the accuracy of the diagnosis;2)the low-resolution problem caused by the accelerated sampling strategy.Although many OCT image denoising methods are developed,in commercial scanners,noiseless OCT images are usually obtained by registering and averaging several OCT images that are repeatedly acquired at the same position of the same sample.Due to the unconscious body jitter or eye movement during the acquisition process,the obtained OCT images used for averaging might not be captured from the exactly same place;consequently,some motion artifacts may appear and some clinically critical information in the averaged OCT images may be lost.Therefore,in the clinical devices,a low sampling rate is often adopted to reduce the influence of unconscious movements.Although this strategy can accelerate the acquisition process,it will introduce the second problem,that is,the resolution of the captured OCT image decreases.To address the above two problems,two algorithms were proposed and implemented based on deep neural network in this paper:(1)Simultaneous denoising and super-resolution of OCT images based on generative adversarial network,dubbed as SDSR.Different from other methods,this algorithm innovatively proposes to couple two separate problems of OCT image,denoising and super-resolution,into a single neural network model.To this end,inspired by previous works,the proposed SDSR consists of an efficient feature extraction module and an iterative up-sampling super-resolution module.By comparing the results of clinical OCT image data set with the existing methods,it can be seen that the SDSR proposed in this paper can effectively remove speckle noise and enhance the resolution of OCT image.Moreover,the higher the scale factor,the more obvious the denoising and super-resolution reconstruction effect of SDSR algorithm compared with other methods.(2)Unsupervised Denoising of OCT images based on generative adversarial network and disentangled representation,termed as DRGAN.The training of supervised learning model requires a large amount of paired data,and in clinical practice,it is very difficult to obtain a sufficient number of fully matched OCT images.Therefore,with the help of generative adversarial network and disentangled representation,an unsupervised OCT image denosing algorithm,called DRGAN,is proposed in this paper.Specifically,DRGAN first disentangles the noisy OCT image into the content space and the noise space by the corresponding encoders,and then uses the clean image generator to reconstruct the denoised OCT image.In addition,the pure noise patches are also included for adversarial learning to further purify the disentanglement of content and noise.Through extensive quantitative and qualitative experiments,it can be seen that DRGAN algorithm is obviously better than some traditional methods in suppressing speckle noise in OCT images,and it is also very competitive compared with some algorithms based on supervised and unsupervised learning.