Abstract
This study aimed to assess the feasibility of an artificial intelligence(AI) model to predict visual acuity(VA) using optical coherence tomography(OCT) in treatment-naïve patients with neovascular age-related macular degeneration (AMD). This retrospective study enrolled 240 patients(240 eyes) with pseudophakic neovascular AMD who received antivascular endothelial growth factor therapy. Each patient underwent 10 visits where they underwent best-corrected visual acuity(BCVA) testing and horizontal OCT scans, yielding 2,400 images. The images were cropped, resized to 224 × 224 pixels, and partitioned at the patient level to avoid data leakage. A pretrained VGG16 CNN was modified for five-class VA classification (< 0.1, 0.1-0.3, 0.3-0.5, 0.5-0.8, ≥ 0.8). The performance was assessed by five-fold cross-validation using the macro-averaged AUC, accuracy, Top-2 accuracy, and binary accuracy (threshold VA = 0.5). The average performance showed a macro-averaged AUC of 0.772, accuracy of 50.3%, Top-2 accuracy of 71.0%, and binary accuracy of 79.6%. For high-confidence predictions (29.2% of the samples), the accuracy improved to 74.1%, with a binary accuracy of 94.2%. ROC analyses demonstrated AUCs of 0.73-0.83 across VA categories, with the strongest discrimination for VA < 0.1 (AUC 0.83). The confusion matrix showed that the VA 0.5-0.8 and ≥ 0.8 categories achieved relatively higher accuracies; however, misclassifications mainly occurred between these adjacent ranges, with frequent bidirectional errors observed. Our pretrained VGG16 showed moderate ability at predicting VA from OCT images in patients with neovascular AMD. While the overall classification was limited, high binary accuracy highlights the potential clinical value of distinguishing good from poor vision.