Fig 1.
Fundus image before (left) and after preprocessing (right).
Fig 2.
The structure of the artificial intelligence model with three input paths which utilizes data on age and sex and ultimately combines the data through element-specific multiplication methods.
Abbreviations: OCT = optical coherence tomography, D = dimension.
Table 1.
Characteristics of patients included in the training/validation and testing sets.
Fig 3.
Representative original fundus photos (A,C) and corresponding heatmaps (B,D). Note that the artificial intelligence prioritizes the region temporal to the fovea.
Table 2.
Results of 5-fold cross validation.
Fig 4.
Ensemble performance of the artificial intelligence model in predicting the incidence of fellow eye neovascularization.
The outcomes from 5-fold cross validation were used to calculate the AUC and F1 score. The precision was 0.562, recall was 0.714, accuracy was 0.667, weighted F1 score was 0.671, AUCROC was 0.675. E1 = examiner 1 (experienced examiner), E2 = examiner 2 (less experienced examiner).
Fig 5.
A representative case illustrating stage 3 fellow eye neovascularization mis-predicted as stage 2 in the synthesized image using artificial intelligence.
A: real optical coherence tomography image at baseline, B: real image when the fellow eye neovascularization was noted, C: artificial intelligence-synthesized image of fellow eye neovascularization. While the synthetic image demonstrates the occurrence of intraretinal fluid, it was unable to accurately predict the development of pigment epithelial detachment.