Latent space autoencoder generative adversarial model for retinal image synthesis and vessel segmentation.

Area of Science:

• Medical image analysis
• Computer vision
• Artificial intelligence in healthcare

Background:

• Diabetic retinopathy (DR) necessitates accurate retinal vessel segmentation for timely diagnosis and monitoring.
• Deep learning models for DR segmentation require large, annotated datasets, which are often scarce.
• Existing methods face challenges due to the lack of sufficient ground truth data for training.

Purpose of the Study:

• To develop a method for generating realistic synthetic retinal fundus images and corresponding segmentation masks.
• To overcome the data scarcity issue in training deep learning models for retinal vessel segmentation.
• To evaluate the effectiveness of synthetic data in improving segmentation performance for diabetic retinopathy detection.

Main Methods:

• Utilized generative adversarial networks (GANs) with a latent space auto-encoder for image synthesis.
• Generated diverse retinal fundus images and binary masks from tubular structured annotations.
• Employed the DRIVE, STARE, and CHASE_DB datasets for data synthesis and validation.
• Trained and tested a UNet model for retinal vessel segmentation using the generated synthetic data.

Main Results:

• The proposed method successfully generated realistic retinal fundus images and accurate binary masks.
• Synthetic data significantly improved the segmentation performance of the UNet model.
• The model trained on synthetic data demonstrated comparable or superior results to models trained on standard datasets.
• Validated the approach's effectiveness across multiple benchmark datasets.

Conclusions:

• Synthetic data generated via GANs is a viable and effective solution for augmenting limited medical imaging datasets.
• This approach holds significant potential for improving the accuracy and generalizability of deep learning models in medical image analysis.
• The method facilitates the development of robust tools for early detection and management of diabetic retinopathy.