Generative Adversarial Network Combined with SE-ResNet and Dilated Inception Block for Segmenting Retinal Vessels.

Area of Science:

• Medical Imaging
• Computer Vision
• Deep Learning

Background:

• Micro vessel segmentation in retinal images is crucial for diagnosing various eye conditions.
• Current segmentation methods face challenges with discontinuity and accuracy.
• Deep learning has significantly advanced image processing efficiency.

Purpose of the Study:

• To develop an accurate generative adversarial network (GAN) for precise retinal micro vessel segmentation.
• To address the issue of discontinuity in current retinal segmentation techniques.
• To improve the accuracy and efficiency of retinal vessel segmentation using deep learning.

Main Methods:

• Proposed an improved GAN, termed SAD-GAN, incorporating SE-ResNet in the generator and a discriminator with dilated inception blocks and attention mechanisms.
• The generator utilizes SE-ResNet to extract global channel information and strengthen key features, while residual structure mitigates gradient disappearance.
• The discriminator enhances feature transmission and deep feature extraction using inception blocks, dilated convolutions, and attention mechanisms for local-global feature integration.

Main Results:

• SAD-GAN demonstrated strong performance on public retinal datasets.
• Achieved ROC_AUC of 0.9813 and PR_AUC of 0.8928 on the DRIVE dataset.
• Achieved ROC_AUC of 0.9839 and PR_AUC of 0.9002 on the CHASE_DB1 dataset.

Conclusions:

• The proposed SAD-GAN significantly enhances retinal vessel segmentation accuracy compared to state-of-the-art methods.
• The integration of generative adversarial networks with deep convolutional neural networks offers a powerful approach for medical image analysis.
• This method holds promise for improved early detection and monitoring of retinal diseases.