Classification of Eye Diseases Using CNN on Fundus Images

Eye diseases represent a critical global health concern, affecting approximately 2.2 billion individuals with visual impairments or blindness and underscoring the urgent need for accessible screening solutions. Early detection is essential for preventing progressive vision loss; however, limited access to eye care significantly delays timely intervention, as witnessed in Indonesia, where more than 8 million cases of blindness and visual impairment have been reported. Fundus imaging detects abnormalities linked to various eye diseases. This system processes fundus images to classify eye diseases. The author trained models using the labeled ODIR (Ocular Disease Intelligent Recognition) dataset. The author's approach incorporates multi-label classification and preprocessing to improve diagnostic accuracy. The author cropped fundus images to reduce background influence and applied Contrast-limited Adaptive Histogram Equalization (CLAHE) for preprocessing. This study evaluates feature extraction methods, including ResNet152, VGG19, and MobileNetV2, to identify the best-performing backbone for automatic eye disease recognition. The system performance was evaluated using the average value of binary accuracy, micro F1-score, AUC, and Cohen's Kappa. According to the experimental findings, the MobileNetV2 model performed best with an F1 score of 88.05%, an AUC of 87.77%, and a Cohen's Kappa of 44.37% while learning at a rate of 0.001. Fine-tuning this model yielded an F1 score of 87.60%, AUC of 88.65%, and Cohen's Kappa of 44.87%.