Efficient Retinal Vessel Segmentation with 78K Parameters

Area of Science:

• Medical Imaging
• Deep Learning
• Ophthalmology

Background:

• Diabetic retinopathy diagnosis relies on retinal vessel segmentation.
• Current deep learning models often face accuracy-complexity trade-offs.
• Efficient segmentation is crucial for clinical settings.

Purpose of the Study:

• To develop a lightweight yet accurate deep learning model for retinal vessel segmentation.
• To address the limitations of existing models in terms of computational complexity and accuracy.
• To facilitate early diagnosis of diabetic retinopathy.

Main Methods:

• Proposed DSAE-Net, a dual-stage network featuring a Parameterized Cascaded W-shaped Architecture.
• Introduced Skeleton Distance Loss (SDL) to manage class imbalance and boundary issues.
• Developed Cross-modal Fusion Attention (CMFA) and Coordinate Attention Gates (CAGs) for enhanced feature refinement.
• Utilized DRIVE, CHASE_DB1, HRF, and STARE datasets for evaluation.

Main Results:

• DSAE-Net achieved superior segmentation accuracy compared to state-of-the-art lightweight models.
• The proposed architecture significantly reduced computational complexity (using only 1% of U-Net parameters).
• The model demonstrated robustness across multiple benchmark datasets.
• SDL effectively handled severe class imbalance inherent in retinal images.

Conclusions:

• DSAE-Net provides an efficient and accurate solution for retinal vessel segmentation.
• The model's low complexity and high performance are suitable for real-time diagnostics.
• This approach supports early detection of diabetic retinopathy in resource-limited environments.