TW-GAN: Topology and width aware GAN for retinal artery/vein classification

Area of Science:

• Medical Image Analysis
• Computer Vision
• Ophthalmology

Background:

• Accurate artery/vein (A/V) classification is crucial for analyzing retinal vascular networks.
• Conventional methods utilize topological and width information, but deep learning methods have not yet integrated these features.
• Existing deep learning approaches for A/V classification lack the exploitation of topological connectivity and vessel width.

Purpose of the Study:

• To propose a novel deep learning framework, the Topology and Width Aware Generative Adversarial Network (TW-GAN), for improved A/V classification.
• To integrate topological connectivity and vessel width information into a deep learning model for retinal vascular network analysis.
• To advance the state-of-the-art in automatic A/V classification using a generative adversarial network.

Main Methods:

• Developed a novel Topology and Width Aware Generative Adversarial Network (TW-GAN).
• Introduced a topology-aware module with a topology ranking discriminator and a topology preserving triplet loss.
• Incorporated a width-aware module to predict vessel width maps for enhanced A/V classification.

Main Results:

• The proposed TW-GAN effectively enhances the topological connectivity of segmented A/V masks.
• Achieved state-of-the-art performance in A/V classification on the AV-DRIVE and HRF datasets.
• Demonstrated the successful integration of topology and width information within a deep learning framework.

Conclusions:

• The TW-GAN framework represents a significant advancement in automatic A/V classification.
• Integrating topological connectivity and vessel width information improves deep learning model performance.
• The method offers a promising approach for quantitative analysis of retinal vascular networks.