부산대학교 도서관

The soaring of diabetes cases in Malaysia has resulted in the appearance of diabetic retinopathy among diabetic patients. Diabetic retinopathy is a chronic eye disease triggered by diabetes, which could worsen eyesight functions and even blindness. Even though the cases were found to be common, medical experts still diagnose the disease manually, which increase the risk of incorrect diagnosis. To overcome this, the preliminary study on severity levels classifications of diabetic retinopathy from fundus images has been conducted by applying a deep learning model. A Convolutional Neural Network (CNN) deep learning model architecture is used to train the dataset, which is DenseNet. Various image pre-processing techniques have been applied to enhance the trained images. Moreover, data augmentation and test-time augmentation (TTA) are implemented in evaluating the training results and lower the overfitting, respectively. Prediction evaluation on the images and the effects of data augmentation and TTA by observing the quadratic weighted kappa values were conducted. Ultimately, a prediction model that is to predict and classify the severity labels of fundus images was developed. The prediction model achieved the quadratic weighted kappa score of 0.9308, with the accuracy of 65% on the Messidor-2 dataset, which were moderately accurate.