A comparative study of a predictive model for iris diseases using machine learning and deep learning methods

This study compares Machine Learning (ML) and Deep Learning (DL) approaches for automated detection of iris diseases — Myopia and Glaucoma — using 7,807 iris images (Healthy, Myopia, Glaucoma) from Kaggle. Two analytical pathways were evaluated under identical conditions: Pathway A used Gabor Filters and 2D Discrete Wavelet Transforms (2D-DWT) with traditional classifiers (SVC, Random Forest, LightGBM, Decision Trees), while Pathway B applied Transfer Learning with CNN_s (XceptionV3, ResNet50, MobileNetV2, EfficientNetB0). All models were assessed via Stratified 5-Fold Cross-Validation across standard classification metrics.
Deep learning models substantially outperformed traditional ML: XceptionV3 achieved the highest accuracy (84.92%), followed by ResNet50 (82.70%) and MobileNetV2 (82.69%). Among ML approaches, 2D-DWT features outperformed Gabor Filters by ~14–15 percentage points, with LightGBM reaching 70.94%. Notably, EfficientNetB0 performed poorly (47.70%), highlighting architecture-dataset sensitivity. These findings affirm CNN_s' superiority for iris disease classification while supporting the continued relevance of ensemble ML methods in data-limited clinical settings.