Retinal disease classification using deep learning techniques

This thesis proposes a deep learning-based classification approach for detecting retinal diseases. The study utilizes the RFMID dataset, which includes multilabel retinal fundus images. In the first phase, a custom-designed Convolutional Neural Network (CNN) was implemented for binary classification, and its performance was compared against various pretrained transfer learning models, including EfficientNetB0, EfficientNetB4, NASNetMobile, VGG16, ResNet152, and DenseNet169. In the second phase, the same transfer learning models were applied to a multiclass classification task to evaluate their broader performance. For binary classification, the two most frequent diseases in the dataset—Diabetic Retinopathy (DR) and Macular Hole (MH)—were selected to address class imbalance and improve model stability. To optimize the decision threshold, ROC curve, Precision-Recall curve, and grid search techniques were employed. Additionally, 5-Fold Cross-Validation was applied to assess the model's generalizability. The results demonstrated that the custom CNN model achieved competitive performance in binary classification compared to transfer learning models, while pretrained networks yielded superior results in multiclass classification. This study highlights the effectiveness of deep learning approaches in retinal image analysis and provides a robust foundation for future research in the field.