Fine-Tuning Deep Convolutional Generative Adversarial Networks (DCGAN) for Synthetic Retinal Fundus Image Generation

Generative adversarial networks (GANs) have shown promise in synthesizing medical images, but their effectiveness is highly sensitive to hyperparameter settings and training strategies. This study investigates the impact of hyperparameter tuning on deep convolutional GANs (DCGANs) for generating synthetic retinal fundus images to support eye disease classification (normal, cataract, glaucoma). We focus on two key research problems: determining suitable hyperparameter configurations for stable training, and assessing whether full-dataset training or class-specific training yields more representative synthetic images. Using Optuna for hyperparameter optimization, the DCGAN model was evaluated with Frechet Inception Distance (FID) and Inception Score (IS). Results show that class-specific training better captures disease-relevant characteristics, while full-dataset training provides more balanced generalization. However, relatively high FID scores and low IS values highlight limitations in image sharpness and diversity, indicating the need for more advanced GAN architectures and larger datasets. This study contributes to the understanding of hyperparameter sensitivity in medical image synthesis and underscores the importance of generalizable approaches for real-world clinical applications.