UNet retinal blood vessel segmentation algorithm based on improved pyramid pooling method and attention mechanism

Area of Science:

• Medical Imaging
• Computer Vision
• Ophthalmology

Background:

• Accurate retinal vessel segmentation is crucial for diagnosing conditions like diabetic retinopathy and hypertension.
• Existing segmentation algorithms face challenges in capturing global context and integrating multi-level features effectively.

Purpose of the Study:

• To enhance retinal blood vessel segmentation performance for improved ophthalmic disease diagnosis.
• To develop a novel segmentation algorithm that integrates pyramid pooling and attention mechanisms.

Main Methods:

• A modified U-Net architecture incorporating a Pyramid Scene Parsing Network (PSP-Net) module for context aggregation.
• Integration of an attention mechanism within the U-Net's skip connections to refine feature fusion.
• Evaluation on the DRIVE and CHASE_DB1 datasets.

Main Results:

• The proposed PSP-UNet algorithm achieved high performance metrics: Sensitivity (0.7814-0.8195), Specificity (0.9810-0.9727), Accuracy (0.9556-0.9590), and AUC (0.9780-0.9784).
• Demonstrated superior performance compared to the standard U-Net and other mainstream retinal vascular segmentation algorithms.
• Effectively enhanced detection of blood vessel pixels and suppressed irrelevant information.

Conclusions:

• The attention-based PSP-UNet algorithm significantly improves retinal blood vessel segmentation accuracy.
• This method offers a promising approach for the automated diagnosis of various ophthalmic diseases.
• The integration of pyramid pooling and attention mechanisms enhances the model's ability to capture global context and relevant features.