A Web-Based Clinical Decision-Support Dashboard for Glaucoma Screening from Retinal Fundus Images: A Two-Phase VGG16 Deep Learning Pipeline with Grad-CAM Interpretability for Ophthalmology Practice

Abstract

Objective:To develop and evaluate a two-phase transfer-learning pipeline using a VGG16 convolutional neural network for automated binary classification of glaucoma from retinal fundus images, designed for deployment in resource-constrained South Asian screening contexts. Methods:A retrospective diagnostic-accuracy study used 705 retinal fundus images from the public ACRIMA dataset. The model employed a VGG16 backbone pre-trained on ImageNet with a custom classification head (global average pooling, two dense ReLU layers with dropout, sigmoid output). Phase-1 trained the head with the backbone frozen (Adam, lr 1 × 10−4, 50 epochs). Phase-2 unfroze the last four convolutional block and fine-tuned at lr 1×10−5 for 25 epochs. Online augmentation, inverse-frequency class weighting, early stopping and learning-rate reduction on plateau were applied. A held-out test set (n = 140, balanced 70/70) was used for final evaluation. A FastAPI inference service and a Next.js clinical dashboard were developed for clinician-facing deployment. Results:The fine-tuned network achieved accuracy 88.57% (95% CI: 82.2–92.8%), sensitivity 92.86% (95% CI: 84.3–96.9%), specificity 84.29% (95% CI: 74.0–91.0%), precision 85.53% (95% CI: 75.9–91.7%), F1-score 89.04%, and AUC-ROC 0.954 (95% CI: 0.918–0.990). Phase-2 fine-tuning raised sensitivity by 7.15 percentage points over Phase-1 (85.71% → 92.86%) and reduced the false-negative rate from 14.3% to 7.1%. Grad-CAM heatmaps consistently localized activation over the optic disc and peripapillary region. Conclusion:A computationally tractable, interpretable VGG16 pipeline reached screening-level sensitivity with a clinician-facing dashboard, providing a reproducible foundation for region-adapted glaucoma screening pending prospective validation on a Pakistani cohort.