ADF-Net: Adaptive Directional Feature Fusion Network for OCTA Vessel Segmentation

Area of Science:

• Biomedical Imaging
• Computer Vision
• Ophthalmology

Background:

• Optical Coherence Tomography Angiography (OCTA) provides high-resolution retinal vascular imaging.
• Accurate segmentation of retinal vessels is vital for vascular analysis and disease assessment.
• Existing methods struggle with complex vascular patterns like thin vessels and bifurcations.

Purpose of the Study:

• To develop an advanced deep learning model for precise retinal vessel segmentation in OCTA images.
• To address limitations in capturing directional continuity and fine-grained details in current segmentation techniques.
• To improve the reliability of retinal vascular analysis for disease detection.

Main Methods:

• Proposed an Adaptive Directional Feature Fusion Network (ADF-Net) utilizing a dual-branch encoder.
• Integrated local feature extraction (convolution) with global context modeling (Swin Transformer).
• Introduced a directional-aware multi-scale module with adaptive aggregation and a direction-wise attention mechanism.

Main Results:

• ADF-Net achieved competitive performance against state-of-the-art methods on OCTA-500 and ROSE-1 datasets.
• Achieved high Dice coefficients (e.g., 91.51% on OCTA-3M) and IoU scores (e.g., 84.42% on OCTA-3M).
• Demonstrated effective preservation of vascular continuity and structural integrity.

Conclusions:

• ADF-Net significantly enhances retinal vessel segmentation accuracy in OCTA images.
• The method effectively handles complex vascular structures, improving diagnostic potential.
• The approach offers a robust tool for retinal vascular analysis and disease assessment.