VDMNet: A Deep Learning Framework with Vessel Dynamic Convolution and Multi-Scale Fusion for Retinal Vessel...

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Retinal vessel segmentation is vital for diagnosing eye and systemic diseases.
• Optical Coherence Tomography Angiography (OCTA) provides detailed retinal microvasculature imaging.
• Current OCTA segmentation methods struggle with noise, class imbalance, and complex vessel structures.

Purpose of the Study:

• To introduce VDMNet, a novel segmentation network for improved OCTA retinal vessel segmentation.
• To address limitations of existing methods in handling noise, class imbalance, and complex vascular patterns.

Main Methods:

• Implemented Fast Multi-Head Self-Attention (FastMHSA) for global and local feature extraction.
• Introduced Vessel Dynamic Convolution (VDConv) for adaptive segmentation of curved and crossing vessels.
• Utilized Multi-Scale Fusion (MSF) for enhanced fine vessel detection and continuity.
• Employed Weighted Asymmetric Focal Tversky Loss (WAFT Loss) to manage class imbalance.

Main Results:

• VDMNet demonstrated effective preservation of tiny vessel edge information.
• Achieved state-of-the-art performance on ROSE-1 and OCTA-3M datasets.
• Showcased superior ability in capturing fine vascular details and overall connectivity.

Conclusions:

• VDMNet offers a robust solution for retinal vessel segmentation from OCTA data.
• The network effectively overcomes limitations of previous segmentation techniques.
• VDMNet's performance highlights its potential for clinical applications in ophthalmic diagnostics.