EXPLORATION OF THE EFFICIENTNET ARCHITECTURE PERFORMANCE IN CLASSIFICATION OF DIABETIC RETINOPATHY ON MULTIDATASETS

• Main Author: Sekti Waskitha Jati, Nur
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/66641
• Institution: Institut Teknologi Bandung

ABSTRACT EXPLORATION OF THE EFFICIENTNET ARCHITECTURE PERFORMANCE IN CLASSIFICATION OF DIABETIC RETINOPATHY ON MULTIDATASETS By NUR SEKTI WASKITHA JATI NIM: 18318023 (Undergraduate Program in Biomedical Engineering) Diabetic retinopathy is one of the complications of diabetes mellitus that can cause blindness at a severe level. The automatic classification model can be used for early detection of the disease so that patients can take early treatment. This model can be applied in places where there is still a shortage of experts to perform manual diagnosis. However, the data owned by each of these places must have their own characteristics. So that the training strategy carried out must be able to overcome the problem of differences in these characteristics, such as small amounts of data, differences in contrast, differences in image resolution, etc. In this study, the optimal strategy is sought for a combination of several datasets or for a specific dataset with certain characteristics. This study uses the APTOS, Messidor-2, and IDRID datasets with the EfficientNetB0 training architecture. In the 5 class classification system for the combination of the three datasets, the optimal strategy obtained is to transfer learning with the initial weight of the APTOS training results. After being given contrast preprocessing, hyperparameter tuning, and resolution optimization, the accuracy obtained is 76.04% and QWK is 0.8873 from an average of 10 fold cross validation. In addition, the strategy used to train per dataset is transfer learning with the initial weight of the training results from the combination of the previous 3 datasets. Contrast preprocessing, hyperparameter tuning, and resolution optimization were also applied so that an average of 10 fold APTOS cross validation resulted in an accuracy of 83.67% and a QWK of 0.9252. For Messidor-2, the accuracy is 72.60% and QWK 0.8251. Meanwhile, for IDRID, the accuracy is 70.91% and QWK is 0.8775. For more applicable purposes, the best results of the 5 class classification in the 2 class classification system were also tested with the accuracy obtained for the combination of 3 datasets, APTOS, Messidor-2, and IDRID, respectively 92.87%; 93.97%; 92.43%; and 93.20%. Meanwhile, the QWK in the classification of 2 classes resulted in successively 0.8498; 0.8763; 0.7969; and 0.8581. Key words : Diabetic retinopathy, APTOS, Messidor-2, IDRID, EfficientNetB0