Structure-aware learnable multi-scale attention for retinal vessel analysis

Area of Science:

• Medical imaging analysis
• Computer vision
• Biomedical engineering

Background:

• Accurate retinal vessel segmentation is vital for diagnosing ocular and systemic diseases.
• Current deep learning models struggle with thin, low-contrast vessels, limiting automated screening.
• Need for improved methods to enhance vessel visibility and structural integrity in retinal images.

Purpose of the Study:

• To develop a novel deep learning module, Learnable Multi-Scale Vesselness Attention (LMSA), to improve retinal vessel segmentation.
• To enhance the ability of deep networks to detect thin, low-contrast vessels.
• To create a more structurally consistent and clinically relevant retinal vessel analysis framework.

Main Methods:

• Proposed the Learnable Multi-Scale Vesselness Attention (LMSA) module, incorporating vesselness priors and directional coherence.
• Integrated LMSA with a U-Net backbone and a multi-task learning (MTL) design for joint supervision of segmentation, boundary, and centerline extraction.
• Validated the framework on public benchmarks (DRIVE, STARE, CHASE_DB1) and an in-house dataset.

Main Results:

• The U-Net+LMSA+MTL model achieved superior performance, outperforming state-of-the-art methods on public datasets.
• Achieved high Dice scores (e.g., 85.39% on DRIVE) and accuracies (e.g., 97.37% on DRIVE).
• Demonstrated strong performance on an in-house dataset (Dice: 82.77%, Accuracy: 95.88%), indicating clinical applicability.

Conclusions:

• The proposed LMSA module effectively embeds vessel prominence and structural priors into deep networks.
• The framework enables more reliable, structurally consistent, and clinically meaningful retinal vessel analysis.
• The method shows significant potential for real-world clinical deployment in automated ocular disease screening.