Fine-Tuning Deep Convolutional Generative Adversarial Networks (Dcgan) Untuk Generasi Data Sintetis Citra Fundus Retina

Abstract

This study aims to develop and analyze the application of Deep Convolutional Generative Adversarial Networks (DCGAN) for generating synthetic retinal fundus images to support the classification of eye diseases. The fundus images used in this research are categorized into three classes: normal, cataract, and glaucoma. The main focus of this study is to investigate the impact of hyperparameter tuning on the performance of DCGAN, and to compare the model’s performance when trained on the entire dataset versus class-specific training. The research consists of several stages, including data preprocessing (resize, crop, normalization), DCGAN architecture design, hyperparameter tuning using Optuna, and model training using two approaches: whole dataset and class specific. The model evaluation is performed using two commonly used GAN metrics: Frechet Inception Distance (FID) and Inception Score (IS). The experimental results show that the best hyperparameter configuration is obtained with a batch size of 64, generator learning rate of 0.0079 (Adam, β₁=0.4748, β₂=0.7258), and discriminator learning rate of 0.0016 (SGD, β₁=0.4192, β₂=0.6781). In the whole dataset training approach, the DCGAN model achieved its best FID score of 113.1952 and IS score of 1.0182. In the class-specific training, the best performance was obtained on the normal class (FID 89.5971; IS 1.0185), while the lowest performance was found in the cataract class (FID 156.0789; IS 1.0126). These results indicate that the class-specific approach allows the model to learn more specific characteristics of each class, while the whole dataset approach enables the model to generalize better to the overall distribution of retinal fundus images.