RESEARCH ON HYBRID ALGORITHMS FOR DIAGNOSING EYE DISEASES

Abstrak

To develop and rigorously evaluate a novel hybrid deep learning framework for simultaneous diagnosis of four critical ocular conditions, more precisely: cataract, diabetic retinopathy, glaucoma, and normal fundus - using a relatively small but balanced dataset of fundus images. The study addresses the challenge of achieving high diagnostic accuracy with limited data through architectural innovation and optimized training protocols. We propose a parallel hybrid convolutional neural network that integrates EfficientNetB3 (for global contextual feature extraction) and DenseNet121 (for local detailed feature extraction). The model processes dual-resolution inputs (300×300 and 224×224 pixels) simultaneously. A novel two-phase training strategy was implemented: Phase 1 (10 epochs) with frozen ImageNet-pre-trained backbones to train only the newly added classification heads, followed by Phase 2 (15 epochs) with selective fine-tuning of upper layers. The model incorporated label smoothing (ε=0.05), L2 regularization, and dropout to combat overfitting. The dataset comprised 3,200 curated fundus images (800 per class), split into training (2,560), validation (320), and test (320) sets. The hybrid model achieved a peak validation accuracy of 92.19% and a test accuracy of 91.87%, significantly outperforming standalone EfficientNetB3 and DenseNet121 models (p<0.001, McNemar's test). Diabetic retinopathy was detected with nearperfect precision (98.75%), while cataract, glaucoma, and normal classes showed robust and balanced performance. The proposed parallel hybrid architecture, combined with a disciplined twophase training regimen, successfully overcomes the limitations of small medical datasets. It effectively leverages complementary feature hierarchies from two state-of-the-art networks, establishing a new benchmark for multi-class ocular disease diagnosis. This work demonstrates that architectural synergy and meticulous training design can yield clinically relevant accuracy without requiring prohibitively large datasets.