OPTIMIZATION OF CONVOLUTIONAL NEURAL NETWORK MODEL MOBILENETV2 PERFORMANCE ON DIABETIC RETINOPATHY SEVERITY CLASSIFICATION

• Main Author: Taufiqurrahman, Shidqie
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/50280
• Institution: Institut Teknologi Bandung

Diabetic retinopathy (DR) is a common complication of diabetes mellitus which which may lead to blindness without early diagnosis and proper treatment. Manual grading of DR severity is known to take a long time and prone to inter-observer variability. Hence, deep learning algorithms has been proposed as one of the automated solutions for DR severity classification. However, most of the successful deep learning models are based on large convolutional neural networks (CNN) architectures, requiring a vast volume of training data and specialized computational resources to train the model. In this study, an optimization strategy that can improve the performance of the CNN model with small architecture in the DR classification has been developed. The strategy that was found to have a positive impact was equalizing the distribution of training data by doing augmentation and sampling of data and using SVM as a classifier. This strategy on the MobileNetV2 architecture (4.2 million parameters) and APTOS 2019 data achieves a performance of 86% accuracy and QWK value of 0.923; which is better or at least comparable with the larger architecture performance on the same dataset. This strategy also can be used on the Messidor dataset with four classes of DR severity, the optimized MobileNetV2 architecture achieves 73% accuracy and QWK value of 0.815. This performance is also better than that achieved by larger architectures. These results indicate that with the right optimization strategy, a CNN model with a small architecture can achieve promising DR classification performance and even outperform the performance of CNN model with a larger architecture.