Artificial Intelligence System Of Typing Retinal Vein Occlusion Based On Fundus Fluorescein Angiography And Its Application

【objective】To explore the application values of artificial intelligence(AI)in fundus fluorescein angiography(FFA)images reading in retinal vein occlusion(RVO),so as to provide AI solutions for the clinical management of patients with RVO.First of all,we established an AI neural network model based on FFA images of RVO using deep learning for automatically identifying non-perfusion areas(NPA)and quantitative analysis,and automatically locating the location of optic disc and macula.Secondly,the established neural network model was used to quantitatively analyze the relationship between the sizes of NPA in patients with central retinal vein occlusion(CRVO)and the occurrence of neovascularization events,so as to provide a reliable and effective predictive index for the clinic.Last,the neural network model was used to locate the specific areas of laser treatment for patients with branch retinal vein occlusion(BRVO),and to guide its targeted retinal photocoagulation accurately.【Methods】This subject is divided into three sections.In section one,we establish the neural network model of FFA images based on deep learning in RVO patients.In section two and three,we explore the value of the model in CRVO and BRVO respectively.1.the 55-degree FFA images of the posterior pole centered on the macula captured by the Heidelberg HRA+OCT multifunctional fundus imaging diagnostic instrument was taken as the research object,and the image segmentation tasks was determined to be NPA,optic disc and macula through pre-experiments.170 FFA images of RVO patients were selected for the construction of NPA segmentation model.After labeling and preprocessing of these images,they were randomly assigned to 136 training sets,17 verification sets and 17 test sets at a ratio of 8:1:1.505 posterior pole FFA images were selected for the construction of the macular and optic disc localization model.After labeling and preprocessing,505 images were randomly assigned to 404 training sets,50 verification sets and 51 test sets at a ratio of 8:1:1.Based on the full convolutional neural network U-Net,an AI image segmentation model for these three tasks is established.The Pixel Accuracy(PA)andIntersection over Union(Io U)were used as evaluation indexes for NPA recognition,and the distance between the center location of optic disc and macular predicted by the model and its actual location was used as evaluation indexes for optic disc and macular location.583 infrared images of normal eyes under same device and same lens were selected to label the optic disc areas,and the sizes of optic disc of FFA images of normal eyes was obtained using statistical analysis.2.The medical records and FFA images of of 343 eyes of 343 CRVO patients diagnosed by clinical examinations during January 2017 to December 2018 were included in this study.The risk of the neovascularization events at least 12 months after the date of onset of the disease was retroactive.Among them,327 patients with327 CRVO eyes were treated with Photoshop for image mosaicking of improved5-field view.The RVO model based on deep learning was applied to quantitatively analyze the sizes of the non-perfusion areas of the 55-degree FFA images of the posterior pole and the mosaic images of improved 5-field view.The ROC curve was used to evaluate the diagnostic value of the NPA sizes of the neovascularization events in CRVO.According to the maximum value of the Youden index,the cut-off point of the neovascularization event was found.3.The FFA images of the posterior pole of 50 BRVO patients diagnosed by clinical examinations during March 2018 to June 2018 were included in this study.The RVO model based on deep learning was applied to automatically identify the NPA of FFA images of the posterior pole of all BRVO patients,and locate their optic disc and macula.Referring to the laser treatment forbidden areas of the PRP in the treatment of PDR,the forbidden areas was determined according to the automatic location of the optic disc and macula.The automatically identified NPA minus the laser treatment forbidden areas is the NPA needed laser treatment that predicted automatically by the AI model.The predicted laser treatment areas was evaluated by experts of fundus ophthalmologists to make sure whether the predicted results were consistent with the actual range.【results】1.The predictive power of the AI model.The average PA of the model for NPA recognition was 0.988,and the average Io U was 0.909.The distance between thepredicted optic disc center and the actual optic disc center was 11.783±5.657 pixels.The distance between the predicted macular center and the actual macular center was7.616(4.472,13.892)pixels.The average prediction time per image was about 3.4seconds.2.The relationship between NPA of CRVO and neovascularization events.In the posterior pole images with 55-degree view,26 out of 343 CRVO eyes had the neovascularization events,with an incidence of 7.58%.The under curve area(AUC)of NPA sizes to indicate the neovascularization events in CRVO was 0.889,the cut-off point was 20.997 DA,the sensitivity was 0.808,and the specificity was 0.946.In the mosaic images of improved 5-field view,23 out of 327 CRVO patients had the neovascularization events,with an incidence of 7.03%.The AUC was 0.921,the cut-off point was 80.834 DA,the sensitivity was 0.783,and the specificity was 0.947.3.The accuracy of the model to predict the laser treatment areas in BRVO.Among the 50 BRVO images predicted by the model,the PA of the model for NPA recognition was 0.911±0.067,and the Io U was 0.875(0.819,0.917).The distance between the predicted optic disc center and the actual optic disc center was15.657±8.061 pixels.The distance between the predicted macular center and the actual macular center was 7.343(0,14.079)pixels.The prediction range of 47 images was consistent with the actual range,accounting for 94%;3 images was inconsistent with the actual range,accounting for 6%.【conclusions】1.The full convolutional neural network U-Net network was used to establish an AI model using deep learning technology based on RVO FFA images for image segmentation and quantification of NPA and location of optic disc and macular center.The test results show that the performance of the model is excellent.2.The combination of AI and FFA images of CRVO patients can provide decision-making basis for classification diagnosis.NPA> 20 DA in the posterior pole under a 55-degree view or NPA> 80 DA under mosaic image of improved 5-field view can be used as a threshold standard for CRVO ischemic typing in the real world.3.The combination of AI and FFA images of BRVO patients can define the specific treatment range of laser photocoagulation of NPA,and provide AI solutionsfor accurate targeted retinal photocoagulation in clinical practice.