Enhancing ResNet with Ghost Weight Normalization For Improved Retina Disease Classification Peningkatan ResNet dengan Ghost Weight Normalization untuk Klasifikasi Penyakit Retina

Abstract

Retinal disease is a dangerous disease. If left untreated, it can cause blurred vision and even cause permanent blindness. Recently, deep learning approaches are widely used to classify medical diseases. A widely used model to classify medical diseases is ResNet. To train the ResNet model, the data used is data obtained from Kaggle with the name Retinal OCT Images (Optical Coherence Tomography) consisting of 4 classes namely choroidal neovascularization (CNV), DRUSEN, diabetic macular edema (DME), and Normal with a total of 83,600 data. The ResNet base model showed accuracy and f1-score of 92%. Modifying the ResNet Base model with the addition of Ghost Weight Normalization (GWN) which aims to provide more weight normalization opportunities shows an increase in accuracy and f1-score to 94%. GWN can also increase the accuracy of CNN Base from 77% to 81%. This improvement shows that GWN can improve the accuracy of Deep learning models with its weight normalization variation technique. Although the training load and training time when using GWN can increase, the accuracy and f1-score of the ResNet model with GWN of 94% can make the chance of misclassification of retinal diseases smaller.