Deep learning based classification analysis for diabetic retinopathy diagnosis

Diabetic retinopathy is an eye disease that occurs in people with diabetes. Early diagnosis of this disease is essential. Artificial intelligence has various applications in the medical field, using advanced algorithms such as big data analytics and deep learning. These systems can be used to support expert doctors' diagnoses or to screen a large population of patients to help with early diagnosis. In this study, an Aptos dataset of 3662 samples of normal, mild, moderate, severe and proliferative fundus eye images and a unique dataset of 7915 images were prepared. The effects of hyperparameters on the classification performance of convolutional neural network models were analyzed using two different datasets. The hyper-parameters evaluated include fixed learning rate, learning rate planning methods, activation functions, L2 regularization, data augmentation techniques and optimization algorithms. Various preprocessing steps were applied to improve the classification success of the images and training was performed using the ResNet34 model with a transfer learning approach. Initially, the model was trained with an SGD optimization algorithm on 3662 Aptos datasets. Then, the model was retrained on 7915 datasets with the Momentum optimization algorithm using data augmentation techniques and L2 regularization, and the results showed an increase of 9.93% in ReLU and 8.99% in Leaky ReLU compared to the initial value. The results showed that the ReLU activation function achieved 0.786 accuracy and F1 score, while Leaky ReLU achieved the highest success with 0.751 accuracy and F1 score. These findings emphasize the importance of hyperparameters in the classification of fundus eye images.