Evaluating the Utility of Synthetic Fundus Images for Medical Imaging AI Models: A Mixed Data Experiment Based on Lesion Classification Performance

Abstract

In recent years, the application of artificial intelligence (AI) in medical imaging has accelerated, yet access to large-scale, balanced clinical datasets remains a critical challenge due to privacy concerns and data sharing limitations. As a potential solution, this study investigates the diagnostic utility of synthetically generated fundus images created using a deep learning-based generative model. (2) Methods: We constructed hybrid datasets by mixing real and synthetic fundus images at eleven different ratios (from 10: 0 to 0: 10) and evaluated the performance of five CNN-based models (Inception V3, Xception, DenseNet121, ResNet50, and MobileNetV2) in classifying four retinal conditions. Key performance metrics-accuracy, precision, recall, and F1-score-were used to quantify classification outcomes. (3) Results: Statistical analyses, including two-way ANOVA and Tukey's HSD post-hoc test, revealed that both model architecture and synthetic data proportion significantly affect performance, with mixing ratios between 4:6 and 5:5 yielding the most stable and accurate results. While models such as DenseNet121 and Xception showed robust performance across a range of mixing ratios, excessive reliance on synthetic data (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\geq 80 \%$</tex>) led to a marked performance decline. (4) Conclusions: These findings suggest that high-quality synthetic fundus images can supplement real data effectively within appropriate limits, offering a viable strategy for data augmentation in AI-based diagnostic systems.

Ähnliche Arbeiten