(DA-U)²Net: double attention U²Net for retinal vessel segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• Retinal morphological changes are key indicators for diagnosing ophthalmic and cardiovascular diseases.
• Manual segmentation of complex retinal vascular structures is inefficient and labor-intensive.

Purpose of the Study:

• To develop an automated and accurate retinal image segmentation network.
• To improve upon the U^2-Net architecture for enhanced feature extraction and noise reduction.

Main Methods:

• Integration of feature channel attention and Convolutional Block Attention Module (CBAM) into the U^2-Net model.
• Development of an Attention-RSU block incorporating spatial and channel attention modules.
• Application of dilated convolution to expand receptive fields and integrate contextual information.

Main Results:

• The proposed network achieved a high accuracy of 98.71% across multiple clinical datasets.
• The attention mechanisms effectively focused on significant features and suppressed noise.
• Dual channel attention reduced detail loss during segmentation.

Conclusions:

• The developed network demonstrates superior performance in retinal image segmentation.
• The model's generalizability suggests potential for application in other medical image segmentation tasks with complex features and scale variations.