Perbandingan Kinerja Arsitektur MobileneTV2 dan MobileneTV3 Dalam Klasifikasi Penyakit Retina pada Citra Optical Coherence Tomography (OCT) Menggunakan Optimizer AdamW dan SGD

Abstract

Retinal diseases are serious visual disorders that can lead to decreased visual function and even blindness. The diagnosis of retinal diseases is generally still performed manually by medical professionals through the examination of Optical Coherence Tomography (OCT) images, a process that requires considerable time, high precision, and is prone to diagnostic errors. Previous studies have mostly employed larger and more complex CNN architectures, with optimization limited to a few commonly used optimizers. This study aims to develop an automatic retinal disease classification model using Convolutional Neural Network (CNN) methods by leveraging the lightweight and efficient MobileNetV2 and MobileNetV3 architectures, enabling faster applications that can be deployed on resource-constrained devices. The architectures evaluated include MobileNetV2, MobileNetV3-Large, and MobileNetV3-Small, along with a comparison of two optimizers, namely AdamW and Stochastic Gradient Descent (SGD). The dataset used consists of 4,000 OCT images divided into four classes: Normal, Choroidal Neovascularization (CNV), Diabetic Macular Edema (DME), and Drusen. The training process was conducted using a transfer learning approach, and model performance was evaluated using accuracy, precision, recall, and F1-score metrics. The experimental results indicate that the combination of the MobileNetV2 architecture with a batch size of 16 and either the AdamW or SGD optimizer achieved the best performance, reaching an accuracy of 85.75%, which is the highest among all tested configurations. These findings highlight the strong potential of lightweight architectures to be developed into fast, accurate, and field-deployable retinal disease diagnostic applications on mobile devices using deep learning.