EFFICIENTNET MODEL OPTIMIZATION FOR FIVE-LEVEL CLASSIFICATION OF DIABETIC RETINOPATHY SEVERITY

• Main Author: Ammar
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/55921
• Institution: Institut Teknologi Bandung

An automatic classification system helps diagnose diabetic retinopathy, especially for applications in areas where there is a shortage of experts for manual diagnosis. Among the various deep learning model architectures, a small architecture is the more ideal choice due to its wider application potential. In a previous study, Taufiqurrahman (2020) has developed a training strategy for MobileNetV2, a small-architecture model. Apart from MobileNetV2, EfficientNet is a group of small-architecture models that have superior image classification performance on various benchmark datasets. In this study, an experimental study was conducted to find an optimization strategy for the EfficientNetB0 model that departed from the strategy developed by Taufiqurrahman. The optimal strategy is obtained by using epoch = 100, batch size = 8, loss = Huber, and optimizer = Adam, without color preprocessing, resampling each class, and applying data augmentation. The optimization results show that EfficientNetB0 trained with the optimal strategy has an accuracy of 80.47% and QWK of 0.9145, with an increase in accuracy of 3.47% and an increase in QWK of 0.0105 from the MobileNetV2 optimal strategy developed in Taufiqurrahman's research. Scaling EfficientNetB0 to EfficientNetB1 with the same input resolution increases total model parameters by 56.83%, but only increases accuracy by 1.84% and QWK by 0.0068, whereas scaling EfficientNetB0 to EfficientNetB2 has a total parameter of 82% greater than EfficientNetB0 only increases accuracy by 1.92% with QWK which is lower than EfficientNetB1. The misclassification of the EfficientNetB0 model that has been trained with an optimal strategy is generally caused by data with clinical features of retinopathy that are not common in its class or not directly related to severity, or data with nonclinical visual features that are unique in its class, such as specific aspect ratios. EfficientNetB1 and EfficientNetB2 share the same misclassification characteristics.