MULTI-CLASSIFICATION OF EYE DISEASES USING A CNN-HARALICK HYBRID FRAMEWORK

Abstract

The detection and classification of eye diseases, including Diabetic Retinopathy, Cataract, and Normal conditions, are critical in medical imaging for early diagnosis and treatment. This study proposes a hybrid CNN-Haralick model, leveraging the lightweight MobileNetV2 CNN architecture for spatial feature extraction and Haralick texture features extraction for texture analysis to enhance the accuracy of eye disease classification. A dual-branch architecture is employed, which fuses features from both the Convolutional Neural Network  and the Haralick-based texture analysis at an early stage. The model is evaluated on a dataset consisting of images from multiple sources. Experimental results show that the hybrid CNN-Haralick model achieves an overall accuracy of 98% on the validation set, outperforming traditional CNN models. The model demonstrates exceptional performance, with a macro average F1-score of 98% for the three classes, and AUC-ROC scores of 100% for each category. The confusion matrix and classification report further validate the model's capability to accurately classify eye diseases, providing reliable decision support for clinicians. Additionally, the model's effectiveness is discussed in comparison with existing works, highlighting its superior performance in terms of both accuracy and computational efficiency.