Predicting Glaucoma Visual Field Progression from Fundus Images with a Weighted Ensemble Method

Early detection of glaucomatous visual field (VF) Progression identification in early stages remains difficult for clinicians to manage, especially in resource-limited settings where access to advanced diagnostic. The availability of tools including optical coherence tomography (OCT) and standard automated perimetry (SAP) remains restricted.This paper proposes a deep ensemble learning framework that uses only widely available color fundus photographs (CFPs) to predict VF progression. The GRAPE dataset uses the PLR3 label to identify substantial VF deterioration while addressing the 5% progressor class imbalance through focal loss and synthetic minority oversampling (SMOTE). The proposed system trains three independent CNN models named EfficientNet-B4 and DenseNet-121 and ConvNeXt-Tiny before using weighted soft-voting to enhance predictive robustness and sensitivity.The proposed ensemble model achieved a classification accuracy of 95%, an F1 score of 0.60, and a recall of 0.75 for the The progression class received superior results compared to the ResNet-50 baseline model which achieved an AUC score of 0.80 in the same dataset.