Noise Reduction And Super-resolution Of Retinal OCT Images Based On Generative Adversarial Network

Optical Coherence Tomography(OCT)has become a powerful technology for clinical diagnosis in ophthalmology due to its advantages of non-invasive and micro-resolution.However,there are two unavoidable problems in clinical diagnosis based on the OCT technology.Firstly,speckle noise exists in the acquired OCT images.Secondly,in clinical circumstance,a low spatial sampling rate is usually used for reducing involuntary movements,which lead to low resolution images.Therefore,suppressing speckle noise and image super-resolution is crucial for improving the quality of OCT images and further image analysis.Many algorithms had been proposed for OCT image denoising and super-resolution,which can be divided into methods based on sparse representation and methods based on deep learning.The former methods introduce artifacts and lose much detailed information,and the latter methods produce images with unclear layer boundaries,smooth details.Recently,generative adversarial networks have achieved excellent results in image translation,and OCT image denoising and super-resolution can also be regarded as image translation tasks.Therefore,we two methods based on generative adversarial network,which are introduced as follows:(1)We propose a supervised method based on attentive back-projection network.We regard image denoising and super-resolution as image translation,under the constraints of the discriminator’s confrontation and the perceptual loss function,the generator can produce high-resolution and clean images.The generator mainly incorporates residual learning,backprojection learning and attention mechanisms.Residual learning is used to extract the deep-level feature of low-resolution images,the back-projection learns the interdependence between low-resolution images and high-resolution images,and the attention mechanism redistributes the attention resources of high-resolution features.The discriminator introduces the idea of Ra GAN,which replaces the absolute probability by calculating the relative true and relative false probability.Experimental results show that the proposed method reconstructed images with clear structure while removing noise and super-resolution.(2)We propose an unsupervised method based on generative adversarial network.The supervised method requires strictly aligned image pairs,and the unsupervised method based on the Cycle-Consistent Adversarial Network(CycleGAN)generates poor image.We propose a new unsupervised method,which treats the image denoising and super-resolution as a two-stage problem,and gradually achieves the task through the CycleGAN network and additional constraint networks.The CycleGAN implements the denoising of the LR image,and the constraint network implements the super-resolution of the LR denoised image.In CycleGAN,the cycle consistency loss and the adversarial loss guide the training of the model.In the constraint network,it is constrained by the reconstructing the loss and the adversarial loss.The experimental results show that the proposed method is superior to the supervised and unsupervised methods in subjective vision,and slightly worse than the supervised method in objective metrics.Finally,the publicly retinal OCT image segmentation software OCTSEG is utilized to automatically segment the layers of the denoised and super-resolved results,and the results further prove the effectiveness of the proposed method.