AA-WGAN: Attention augmented Wasserstein generative adversarial network with application to fundus retinal vessel...

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate segmentation of retinal blood vessels is crucial for diagnosing various eye diseases.
• Complex vascular structures and small vessels present significant challenges for automated segmentation algorithms.

Purpose of the Study:

• To propose a novel Attention Augmented Wasserstein Generative Adversarial Network (AA-WGAN) for enhanced fundus retinal vessel segmentation.
• To improve the segmentation of intricate and tiny vascular structures in retinal images.

Main Methods:

• A U-shaped network generator incorporating attention augmented convolution and a squeeze-excitation module was developed.
• The Wasserstein generative adversarial network backbone utilized a gradient penalty method to stabilize training.
• The model was evaluated on three public datasets: DRIVE, STARE, and CHASE_DB1.

Main Results:

• The proposed AA-WGAN achieved high accuracy rates of 96.51% (DRIVE), 97.19% (STARE), and 96.94% (CHASE_DB1).
• The attention augmented convolution effectively captured long-range dependencies for highlighting regions of interest.
• Ablation studies confirmed the effectiveness of individual components and the model's generalization ability.

Conclusions:

• The AA-WGAN is a competitive and effective model for retinal vessel segmentation, particularly for challenging cases with small or complex vessels.
• The integration of attention mechanisms and generative adversarial networks offers a promising approach for medical image analysis.
• The model demonstrates robust performance and generalization across different datasets.