Research On The Diagnosis Of Diabetic Macular Edema Based On Artificial Intelligence

PurposeTo develop deep learning algorithms for identification of diabetic macular edema(DME)from color fundus photographs(CFP),which are different from previous published algorithms in that the grading of DME in CFPs were based on evaluation of optical coherence tomography(OCT)images from the same patients.To develop deep learning algorithm for identification of different morphological types of DME from OCT images.MethodsRetrospective-analysis of 387 OCT volumes and 297 color fundus photographs of diabetes mellitus patients.Each CFP is paired with the OCT images obtained from the same impaired eye.According to morphological feature and average retinal thickness of the central lmm-diameter circle of Early Treatment Diabetic Retinopathy Study(ETDRS)grid from OCT,all OCT images and their paired CFPs were categorized into central involved diabetic macular edema(CIDME)and n-CIDME,DME and n-DME.OCT images were also labeled with DME types.All these images were randomly divided into training set and validation set.Deep learning models were developed to identify and grade macular edema from CFPs and OCT images.Sensitivity and specificity in identifying DME and CIDME were compared between CFP-based algorithm and OCT-based algorithm,and results were analyzed based on different DME morphological types and existence of hard exudate.Each OCT image were given a label of DME morphological type,which includes diffuse retinal thickening(DRT),cystoid macular edema(CME),serous retinal detachment(SRD)and posterior hyaloid traction(PHT).Labeled OCT images were used to train deep learning algorithm to automatically identify DME type of unclassified OCT images.ResultsFor the detection of DME,the CFP-based model achieved a sensitivity of 0.73 and specificity of 0.93,while the sensitivity and specificity of OCT-based model were 0.90 and 0.97,respectively.In terms of the detection of CIDME,the CFP-based model had a sensitivity of 0.72 and specificity of 0.74,compared to the sensitivity of 0.89 and specificity of 0.98 of OCT-based model.Posterior hyaloid traction(PHT)and serous retinal detachment(SRD)were relatively easy to be detected from both OCT images and fundus photographs in comparison to other DME types.The deep learning models trained by OCT images with labels on DME types was able to identify DRT with a sensitivity of 0.92 and a specificity of 0.94,CME with a sensitivity of 0.95 and a specificity of 0.98,SRD with a sensitivity of 0.92 and a specificity of 0.99,PHT with a sensitivity of 0.92 and a specificity of 0.94,normal with a sensitivity of 0.92 and a specificity of 0.94.ConclusionDeep learning algorithm,trained by CFPs labeled as CIDME/n-CIDME,and DME/n-DME according to features of paired OCT images,is capable of detecting DME and CIDME from CFPs,although OCT-based model performs better in both DME and CIDME detection.Deep learning algorithm can differentiate DME types in OCT images with high sensitivity and specificity.