Artery-vein segmentation in fundus images using a fully convolutional network.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Retinal vessel caliber analysis is crucial for diagnosing ocular diseases, coronary heart disease, and stroke.
• Manual differentiation of arterioles and venules in fundus images is time-consuming and prone to variability.
• Automating this process is essential for large-scale epidemiological studies.

Purpose of the Study:

• To assess the potential of fully convolutional networks (FCNs) for automated retinal artery-vein (A/V) discrimination.
• To develop and evaluate a novel deep learning (DL) architecture for simultaneous vessel extraction and A/V discrimination.
• To improve the automation and accuracy of retinal vessel analysis.

Main Methods:

• A U-Net semantic segmentation architecture, a type of FCN, was applied to discriminate arteries and veins in fundus images.
• The DL model was trained and tested on publicly available datasets: DRIVE and HRF.
• Performance was evaluated based on accuracy metrics for artery and vein identification.

Main Results:

• The FCN model achieved high accuracies: 94.42% for arteries and 94.11% for veins on the DRIVE dataset (vessels >2 pixels wide).
• This represents a 25% reduction in error compared to previous state-of-the-art methods.
• On the HRF dataset, the model achieved 96.98% accuracy on all identified centerline pixels.

Conclusions:

• Deep learning, specifically the U-Net architecture, offers a powerful and accurate solution for automated retinal artery-vein discrimination.
• The developed model significantly outperforms existing methods, reducing manual labor and variability in vessel analysis.
• This advancement has the potential to enhance the efficiency and reliability of diagnosing systemic diseases through retinal imaging.