Electronic Theses and Dissertation

Gangguan penglihatan akibat penyakit mata dapat menyebabkan kebutaan apabila tidak terdeteksi sejak dini. Penyakit seperti Katarak, Glaukoma, dan Retinopati Diabetik dapat dianalisis melalui citra fundus retina. Penelitian ini bertujuan menganalisis dan membandingkan performa dua arsitektur deep learning lightweight, yaitu MobileNetV2 dan EfficientNetB0, dalam klasifikasi penyakit mata multi-kelas menggunakan citra fundus. Pendekatan transfer learning diterapkan dengan membekukan bobot base model dan melatih arsitektur head model yang terdiri dari lapisan Global Average Pooling, Dense 256 neuron dengan aktivasi ReLU, Dropout 0,3, dan output layer Softmax. Dataset yang digunakan merupakan dataset publik citra fundus berjumlah sekitar 4.217 citra yang terbagi ke dalam empat kelas, yaitu Normal, Katarak, Glaukoma, dan Retinopati Diabetik. Evaluasi performa dilakukan menggunakan metrik accuracy, precision, recall, dan F1-score. Hasil pengujian menunjukkan bahwa EfficientNetB0 menghasilkan performa yang lebih unggul dibandingkan MobileNetV2 pada seluruh metrik evaluasi, dengan akurasi sebesar 88,82%, precision 0,8943, recall 0,8882, dan F1-score 0,8887, dibandingkan MobileNetV2 yang memperoleh akurasi 83,46%, precision 0,8461, recall 0,8346, dan F1-score 0,8371. Berdasarkan hasil tersebut, EfficientNetB0 ditetapkan sebagai arsitektur lightweight yang memberikan performa terbaik dalam klasifikasi multi-kelas penyakit mata berbasis citra fundus pada penelitian ini.

Kata kunci: Citra Fundus, Deep Learning, Lightweight Model, Klasifikasi Multi-Kelas, MobileNetV2, EfficientNetB0

Visual impairment caused by ocular diseases can lead to blindness if not detected at an early stage. Diseases such as cataract, glaucoma, and diabetic retinopathy can be analyzed through retinal fundus images. This study aims to analyze and compare the performance of two lightweight deep learning architectures, namely MobileNetV2 and EfficientNetB0, for multi-class ocular disease classification using fundus images. A transfer learning approach is applied by freezing the base model weights and training a head model architecture consisting of a Global Average Pooling layer, a Danse layer with 256 neurons and ReLU activation, a Dropout layer with a rate of 0.3, and a Softmax output layer. The dataset used is a public fundus image dataset comprising 4,217 images categorized into four classes: Normal, Cataract, Glaucoma, and Diabetic Retinopathy. Model performance is evaluated using accuracy, precision, recall, and F1-score metrics. The experimental result show that EfficientNetB0 outperforms MobileNetV2 across all evaluation metrics, achieving an accuracy 0f 88.82%, precision of 0.8943, recall of 0.8882, and F1-score of 0.8887, compared to MobileNetV2 which obtained an accuracy of 83.46%, precision of 0.8461, recall of 0.8346, and F1-score of. 8371. Based on these results, EfficientNetB0 is established as the lightweight architecture that delivers the best performance for multi-class ocular disease classification based on fundus images in this study. Keywords: Fundus Image, Deep Learning, Lightweight Model, Multi-class classification, MobileNetV2, EfficientNetB0