Klasifikasi Severity Level Diabetic Macular Edema Berbasis ResNet-50

Abstract

Diabetes is one of the most common diseases people suffer from today, and it can lead to complications such as blindness, heart disease, and kidney failure. The condition of blindness caused by this disease is known as diabetic retinopathy (DR). An ophthalmologist will use a fundus camera to examine the retina, looking for several clinical features, such as microaneurysms (MA), hemorrhages (HM), cotton-wool spots (CWS), and exudates. Based on these clinical symptoms, clinicians then determined the patient's level of diabetic macular edema (DME) severity. Although several studies have applied CNN-based architectures for diabetic retinopathy detection, limited attention has been given to the impact of dataset imbalance handling on DME severity classification, particularly using ResNet-50. This study highlights the significant impact of extensive data augmentation on classification performance in imbalanced DME datasets. Evaluate performance using the accuracy, precision, and recall metrics. We used the IDRiD dataset, which consists of 516 images split into a training set of 413 and a test set of 103. IDRiD divides the dataset into three classes, namely normal, moderate DME, and severe DME. In the preprocessing stage, we enhanced contrast using CLAHE and resized the images to 224x224 pixels. To address the imbalance, we applied 11 data augmentation methods. We experimented by comparing the performance of two models: one with and one without dataset augmentation. Based on the test results, the best performance was obtained with the model that included dataset augmentation, achieving an accuracy of 0.5961, a precision of 0.63, and a recall of 0.61, while the baseline model (without dataset augmentation) gained 0.4553, 0.36, and 0.34 for the accuracy, precision, and recall, respectively.