THE EFFECT OF CLAHE-BASED CITRA QUALITY IMPROVEMENT ON CLASSIFICATION OF DIABETIC RETINOPATHY SEVERITY LEVELS USING RESNET50

• Main Author: Ode Ansyarullah S. Sagala, La
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/86696
• Institution: Institut Teknologi Bandung

Diabetic Retinopathy (DR) is a microvascular complication of the retina caused by Diabetes Mellitus (DM), which can potentially lead to vision impairment and even blindness if not detected and treated early. Early diagnosis of DR is typically performed through fundus image analysis by ophthalmologists. However, this manual method is time-consuming and requires significant expertise in fundus image analysis. As an alternative, Computer-Aided Detection (CAD) technology offers a more efficient solution to accelerate DR diagnosis. Deep Learning (DL) approaches have been widely employed to automatically classify DR severity levels, yielding promising results. However, challenges such as variations in the size and width of fundus images that can impact model performance, the need for large datasets, and the significant computational resources required for optimal model training remain. This study focuses on classifying DR severity levels using the ResNet50 architecture. The dataset used is APTOS2019, consisting of 3,662 fundus images. The dataset was divided into training data (90%) and testing data (10%). The preprocessing scenario applied included processing the green channel using CLAHE (Contrast Limited Adaptive Histogram Equalization), followed by resampling to balance the distribution of training data, as well as data augmentation to increase dataset variability. Furthermore, this study employed an ensemble method, combining several classifiers such as SVM, Random Forest, and Logistic Regression. The evaluation results showed an average accuracy of 83%, precision of 69%, recall of 66%, and f1-score of 67% using a 10-fold cross- validation scheme. The best performance was achieved with an accuracy of 85%, precision of 72%, and recall and f1-score of 71% each. In terms of AUROC, the macro-average AUROC achieved was 0.96, with ROC values for each class as follows: Normal at 1.00, Mild at 0.96, Moderate at 0.95, Severe at 0.95, and Proliferative at 0.91. The results of this study demonstrate a significant improvement in the model's performance for classifying DR severity levels, indicating that the combined approach of preprocessing, augmentation, and ensemble methods can be an effective solution to enhance classification accuracy.