Disease-specific data augmentation enhances deep learning classification of age-related macular degeneration,...

To determine which data augmentation technique yields the best performance for deep learning models in classifying age-related macular degeneration (AMD), diabetic retinopathy (DR), glaucoma, and normal fundus images.

This study employed an in silico experimental study design. Six data augmentation techniques: Colour Jitter, Contrast-Limited Adaptive Histogram Equalisation (CLAHE), Rotation, Translation, Gaussian Noise, and Poisson Noise were evaluated using controlled experiments with an EfficientNet-B0 model on a balanced dataset of 1,200 fundus photographs, 250 cases each for AMD, DR and glaucoma, and 450 normal fundus images curated from four main publicly available databases. The experiments were conducted in four phases: baseline, single augmentations, combined augmentations, and the impact of augmented dataset volume. Evaluation metrics and visualisations were computed with Python-based statistical and visualisation libraries.

The results from this study show that data augmentation consistently increased the area under the curve (AUC) from 96.55% to 97.23% and accuracy from 85.83% (baseline) to 89.58%. The results indicate that augmentation effectiveness is disease-specific: Rotation and Colour Jitter yielded the highest sensitivity for AMD (99%), CLAHE maximised sensitivity for Diabetic Retinopathy (96%), and Translation was most effective for Glaucoma (83%). While single augmentations provided descriptive clinical improvements, the comprehensive combination of photometric, geometric, and noise augmentations yielded the best overall performance and achieved a statistically significant improvement over the baseline (Mean bootstrapped AUC = 0.9800, 95% CI: 0.9678, 0.9895; p= 0.0050).

Data augmentation effectiveness is disease-dependent; specific pathologies respond better to distinct augmentation techniques due to different retinal biomarkers.