Development of Deep Learning Classification Model for Diabetic Retinopathy Detection and Grading

• Main Author: Nurul Mirza Afiqah, Tajudin
• Format: Thesis
• Language: English; English
• Published: UNIMAS 2023
• Subjects: QA75 Electronic computers. Computer science
• Online Access: http://ir.unimas.my/id/eprint/43167/5/retricted%20tesis_Nurul%20Mirza.pdf; http://ir.unimas.my/id/eprint/43167/10/Nurul%20Mirza%20Afiqah%20ft.pdf; http://ir.unimas.my/id/eprint/43167/
• Institution: Universiti Malaysia Sarawak

The increasing number of cases of undiagnosed diabetic patients worldwide has been worrisome as more underlying diseases or side effects of diabetes can go undetected, such as Diabetic Retinopathy (DR). The lack of medical assistance in remote areas is a huge problem as it is a long travel distance to the city and comes with a high cost. Hence, studies have been working towards automating medical diagnosis by incorporating Artificial Intelligence (AI) in their system. The rapid growth of technologies and AI has led to the development of Deep Learning (DL), in which its algorithms are stacked in a hierarchy of increasing complexity and abstraction. Thus, by employing DL in the study, automated DR screening will ease the process of diagnosing DR and grading the severity level of the disease. The EyePACS data used in the study are from a competition organized in Kaggle, which consists of 35,126 training and 1,794 testing images that have been graded to their severity level; 0: No DR; 1: mild Non-Proliferative DR(NPDR); 2: Moderate NPDR; 3: Severe NPDR; and 4: Proliferative DR (PDR). The datasets are fed to the Convolutional Neural Network (CNN) in which the model was trained to learn to classify the features of each severity level of DR. Several CNNs are compared in terms of their performance to grade DR, which is AlexNet, ResNet-18, GoogLeNet, Inception-V3, MobileNetV2, and VGG-16. Then, Inception-V3 is picked to be further investigated under different configurations and parameter settings. The features that the model learnt for each class are explored to understand the process behind CNN layers. The final testing accuracy achieved by the model is 80.10%, with sensitivity of 0.4248 and specificity of 0.8860.