Application Of The Generative Adversarial Networks To The Restoration Of Blurred Images Of Optical Coherence Tomography

Purpose: Restoration of blur optical coherence tomography images due to media opacity using the generation adversarial network(GANs)technique.Methods: This study collected 272 eligible subjects(510 eyes)from December 2018 to October 2020 at JOINT SHANTOU INTERNATIONAL EYE CENTER(JSIEC).All subjects were scanned three times in the macular cube 512 * 128 mm mode of Cirrus OCT,and the images with the highest signal strength(SS)were selected for preservation.This research data altogether divides into 4 data sets: 1.Training set: 210 subjects(420 eyes)with transparent dioptric media were enrolled.The clear images(Y)were obtained by direct scanning the macula,and the blurred images(X)were obtained by fuzzy processing of clear images with a machine algorithm,the number of both picture X and Y is 3,702.2.test set neural density filter(NDF): 28 subjects(56 eyes)were included.Y and X were composed of images before and after scanning with filter(optical density0.69),the number of both picture X and Y is 100.3.The test set cataract,34 subjects(34 eyes)who had completed the cataract surgery with cataract only.Y was from OCT images more than 1 month after surgery and X was from OCT images before surgery,the number of both picture X and Y is 504.The test set generated graph: 62 subjects(90 eyes)in the filter group and cataract group were included.Y was the same as the two groups.X was obtained by fuzzy processing y using a machine algorithm,the number of both picture X and Y is 770.The model is trained by pairing Y and X in the training set,and the model is tested by X in the three test sets,and the corresponding Y is used as the standard picture,the performance of GANs model for restoration of blurred images is verified by image evaluation index Peak signal-to-noise ratio(PSNR)and structural similarity(SSIM).The Wilcoxon test was used to test the statistical differences between before and after the image restoration.The set P < 0.05 had statistical differences.Results: In test set NDF,before reduction,PSNR was 18.3650±0.4392,SSIM was0.8528±0.0160,after reduction,PSNR was 19.9398±0.2916,SSIM was 0.9986±0.0019.In test set cataract,before reduction PSNR was 16.6527±0.9893,SSIM was 0.9247±0.0700,after reduction,PSNR was 16.9109±0.2597,SSIM was0.9730±0.0175.In test set generated graph,before reduction,PSNR was 18.3263±0.5458,the SSIM was 0.8555±0.0207,after reduction,PSNR was20.8316±0.4086,SSIM was 0.9933±0.0063.The values of PSNR and SSIM before restoration were higher than those after restoration(p < 0.01,Wilcoxon test).Among the three groups,the result of generating test set was the best,the filter group was the second,and the cataract group was the worst,the difference was statistically significant(p < 0.01,Kruskal-Wallis test).Conclusions: GANs can be suggested to restore blurred OCT images caused by opacities of media opacity.Further research could collect enough real pre-and post-operative images of the cataract to use the pix2 pix model for paired training,and increase the types of disease in the training images,to enhance the feasibility of clinical application.