Enhancing OCTA Image Classification Using Superpixel-Derived Geometric and Texture Features

Abstract

Optical Coherence Tomography Angiography (OCTA) is an important imaging technique for diagnosing and monitoring retinal diseases. However, the accurate classification of OCTA images remains challenging due to the complexity of vascular structures and imaging variability. This study introduces a novel approach that enhances OCTA image classification for Diabetic Retinopathy (DR) and Myopia by leveraging superpixel-derived geometric and texture features (Mean Intensity, Area, Perimeter, Compactness, Eccentricity, Contrast and Entropy). The proposed method is evaluated using the FAZID dataset, which contains 304 Superficial Vascular Plexus (SVP) OCTA images classified into Diabetic (107), Myopic (109) and Normal (88) cases. Six machine learning models—Decision Tree, Random Forest, XGBoost, Extra Trees, LightGBM and CatBoost—were tested to assess classification performance. Experimental results indicate that boosting-based classifiers, such as XGBoost, LightGBM and CatBoost, achieved 100% classification performance in terms of accuracy, precision, recall, F1-score and MCC. Among bagging classifiers, Random Forest achieved 95.56% accuracy, 95.67% precision, 95.37% recall, 95.50% F1-score and 93.32% MCC, while Extra Trees obtained 95.84% accuracy, 96.08% precision, 95.58% recall, 95.77% F1-score and 93.76% MCC. Additionally, the Decision Tree classifier achieved 100% accuracy across all metrics. This study highlights the impact of superpixel-based feature representation combined with machine learning techniques, offering a robust solution for automated OCTA image analysis in ophthalmology.