Dynamic Statistical Attention-based lightweight model for Retinal Vessel Segmentation: DyStA-RetNet.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Accurate retinal vascular segmentation is crucial for diagnosing and treating eye diseases.
• Existing deep learning models struggle with complex vessel structures, false positives, and computational demands.
• Deployment in resource-constrained environments is hindered by model complexity.

Purpose of the Study:

• To develop an attention-based, computationally efficient architecture for improved retinal vessel segmentation.
• To address limitations of existing models in accuracy and deployability.

Main Methods:

• Proposed DyStA-RetNet: a shallow CNN encoder-decoder architecture.
• Incorporated partial decoder for high-level semantic transfer and a separate branch for low-level information.
• Utilized multi-scale dynamic attention and statistical spatial attention blocks for enhanced feature learning.

Main Results:

• DyStA-RetNet demonstrated superior segmentation performance across four benchmark datasets (DRIVE, STARE, CHASEDB, HRF).
• Achieved significantly fewer trainable parameters (37.19K) and GFLOPS (0.75).
• Showcased adaptability for clinical applications.

Conclusions:

• The lightweight DyStA-RetNet efficiently extracts complex retinal vascular components.
• The model is computationally efficient and suitable for resource-constrained environments.
• Enables improved diagnostic capabilities for retinal diseases.