Synthesizing realistic high-resolution retina image by style-based generative adversarial network and its utilization

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning-based realistic image synthesis is crucial for developing AI diagnostic tools and protecting patient privacy.
• Training generative adversarial networks (GANs) for image synthesis requires substantial data, posing challenges for diverse image features.

Purpose of the Study:

• To synthesize retinal images that are indistinguishable from real ones.
• To evaluate the efficacy of synthesized retinal images for augmenting imbalanced datasets in disease classification.
• To assess the potential of synthesized images in improving computer-aided diagnosis systems.

Main Methods:

• Generative Adversarial Network (GAN) for realistic retinal image synthesis.
• Image Turing tests involving human experts and quantitative image analysis.
• Deep learning-based classification performance evaluation on augmented datasets.

Main Results:

• Synthesized retinal images demonstrated high realism, validated by expert qualitative analysis and quantitative vessel metrics (0.43% difference in amount, 1.5% in SNR).
• Turing tests showed moderate performance in distinguishing real from synthetic images (accuracy 54.0%, sensitivity 71.1%, specificity 36.9%).
• Augmenting datasets with synthesized images significantly improved deep learning classification performance across various imbalance ratios.

Conclusions:

• Realistic retinal images can be successfully generated using GANs with minimal differences from real images.
• Synthesized retinal images hold significant potential for practical applications in medical imaging, particularly for augmenting training datasets and enhancing AI diagnostic capabilities.
• This approach offers a viable solution for data scarcity and privacy concerns in developing AI-driven medical diagnosis systems.