Dual-path multi-scale context dense aggregation network for retinal vessel segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Deep Learning

Background:

• Deep learning shows promise for blood vessel segmentation but struggles with limited data, context, and microvascular detail loss.
• Existing methods often fail to capture fine vascular structures and context effectively.

Purpose of the Study:

• To propose a dual-path deep learning framework for enhanced blood vessel segmentation.
• To address challenges like small sample sizes, context neglect, and microvascular detail loss in current methods.

Main Methods:

• A dual-path framework divides fundus images into multi-scale concentric patches.
• The Multi-scale Context Dense Aggregation Network (MCDAU-Net) incorporates Cascaded Dilated Spatial Pyramid Pooling (CDSPP) and InceptionConv (IConv) modules.
• A Multi-scale Adaptive Feature Aggregation (MAFA) module and a fusion module combine multi-scale features for fine segmentation.

Main Results:

• The MCDAU-Net framework demonstrated significant improvements on the DRIVE, CHASE-DB1, and STARE datasets.
• Achieved a mean increase of 7.9% in Sensitivity (Se) and 4.7% in F1 score compared to state-of-the-art methods.
• Successfully enhanced segmentation of low-contrast vessels and preserved microvascular continuity.

Conclusions:

• The proposed dual-path deep learning framework effectively overcomes limitations in blood vessel segmentation.
• The novel architecture and modules significantly advance the state-of-the-art in retinal vascular segmentation.
• The method offers a robust solution for accurate and detailed blood vessel analysis in medical imaging.