Implementasi Model Convolutional Neural Network Efficientnetv2b3 pada Citra Fundus untuk Klasifikasi Tingkat Keparahan Diabetes Retinopati

Abstract

Diabetic retinopathy (DR) is a serious complication of diabetes mellitus that can leadto permanent blindness if not detected and treated early. This study implements a diabetic retinopathy classification system using a deep learning approach with EfficientNetV2B3 architecture and transfer learning. The dataset used is a combination of three main sources: APTOS 2019, IDRID, and MESSIDOR fundus images classified into five severity categories: No DR (Grade 0), Mild (Grade 1), Moderate (Grade 2), Severe (Grade 3), and Proliferative (Grade 4). The model was trained using preprocessing techniques including resizing to 224×224 pixels, normalization, and data augmentation with RandomFlip and RandomRotation. Transfer learning was applied with pre-trained weights ImageNet and a custom head for five-class classification. Experiments were conducted with various epoch configurations (50, 70, 80, 100, 200) using batch size 64. The best results were obtained with the 200-epoch configuration achieving 92.71% validation accuracy. The model achieved balanced F1-scores across all classes with the highest performance on No_DR class (F1-score 0. 9801), Proliferative (F1-score 0. 9223), and Severe (F1- score 0.8735). The model was then converted to TensorFlow Lite format for mobile application. This system has the potential to assist diabetic retinopathy screening processes in resource-limited healthcare facilities and provide real contributions in early detection to prevent vision loss.