Klasifikasi Diabetic Retinopathy Menggunakan Machine Learning Dengan Metode Convolutional Neural Network Model Arsitektur EfficientNetB3

Abstract

Diabetic retinopathy is a serious complication of diabetes mellitus that can lead to blindness if not detected and treated early. This study aims to develop a Convolutional Neural Network (CNN)-based machine learning model with EfficientNetB3 architecture to categorize and determine the severity of diabetic retinopathy through analysis of retinal OCT fundus images. The research methodology includes downloading a dataset from Kaggle consisting of 10,000 retinal images divided into five classes: Healthy, Mild Diabetic Retinopathy, Moderate Diabetic Retinopathy, Severe Diabetic Retinopathy, and Proliferative Diabetic Retinopathy, each class containing 2000 images. The data preprocessing process was carried out by resizing the image from 256x256 pixels to 224x224 pixels and applying data augmentation in the form of brightness, rotation and flip to increase the variation of the dataset. The EfficientNetB3 model was implemented using transfer learning with the TensorFlow/Keras framework, with batch normalization, dense, and dropout layers added, and then compiled using the Adam optimizer, binary crossentropy loss function, and accuracy metrics. The dataset was divided into 90% training data and 10% validation data, with the application of data augmentation techniques, weighted loss, cosine annealing, and test time augmentation (TTA) to address class imbalance and improve model generalization. The results showed that the model successfully achieved a validation accuracy of 95.8% with an average F1-score of 0.9570. The best performance was shown in the classification of healthy eyes (Healthy) with an F1- score of 0.98, while the Mild and Moderate classes showed slightly lower performance due to similar visual characteristics. Evaluation through confusion matrix and classification report confirmed that the model has high precision and recall in almost all classes.