IDENTIFIKASI PENYAKIT MATA PADA CITRA FUNDUS MENGGUNAKAN MODEL HIBRIDA TRANSFER LEARNING–XGBOOST

Abstract

Eye diseases constitute a major health concern that may lead to visual impairment or blindness if not detected and treated at an early stage. Manual analysis of fundus images requires considerable time, accuracy, and ophthalmological expertise; therefore, a computational method capable of supporting automated and objective classification is needed. This study aims to develop a hybrid transfer learning and XGBoost model for classifying fundus images into four ocular condition categories. The methodology involved applying Contrast Limited Adaptive Histogram Equalization (CLAHE), resizing the images to 224 × 224 pixels, performing retina-specific augmentation on the training data, and stratifying the dataset into training, validation, and testing sets. An ImageNet-pretrained EfficientNetV2S model was employed through transfer learning and fine-tuning to extract image features. The resulting 256-dimensional embeddings, model output probabilities, and statistical image features were subsequently combined, standardized, and classified using XGBoost optimized through five-fold cross-validation. The dataset comprised 4,217 fundus images distributed across four classes: 1,038 cataract images, 1,098 diabetic retinopathy images, 1,007 glaucoma images, and 1,074 normal images, including 2,949 training, 631 validation, and 637 testing images. Evaluation on the 637 test images demonstrated that the proposed transfer learning–XGBoost model achieved an accuracy of 92.78%, macro-precision of 0.93, macro-recall of 0.93, macro F1-score of 0.9265, and one-vs-rest ROC-AUC of 0.9871. The proposed model outperformed the standalone EfficientNetV2S model, which achieved an accuracy of 90.27%, by 2.51 percentage points, with class-specific F1-scores of 0.96 for cataract, 0.99 for diabetic retinopathy, 0.87 for glaucoma, and 0.88 for normal. These findings demonstrate that integrating transfer learning with XGBoost effectively improves eye disease classification and has the potential to support early identification using fundus images, although it is not intended to replace clinical diagnosis by ophthalmologists.