Multi-Task Siamese Network for Retinal Artery/Vein Separation via Deep Convolution Along Vessel.

Area of Science:

• Medical image analysis
• Ophthalmology
• Computer vision

Background:

• Retinal artery-vein (A/V) separation is crucial for analyzing vascular diseases.
• Existing graph-based methods treat vessel disentanglement and A/V classification independently, often leading to suboptimal results.
• Ad hoc rules and hand-crafted features limit the performance of current A/V separation techniques.

Purpose of the Study:

• To develop a joint learning framework for vascular tree disentanglement and A/V type classification.
• To improve the accuracy and robustness of retinal A/V separation.
• To overcome the limitations of independent task processing in existing methods.

Main Methods:

• Proposed a multi-task Siamese network for joint learning of A/V separation tasks.
• Introduced Convolution Along Vessel (CAV) to extract visual and geometric features.
• Trained the network for A/V type classification and segment similarity estimation to disentangle vasculature.

Main Results:

• Achieved high accuracy rates of 94.7% (DRIVE), 96.9% (INSPIRE), and 94.5% (WIDE).
• Demonstrated superior performance compared to recent state-of-the-art methods.
• Showcased the effectiveness of joint learning for robust A/V separation.

Conclusions:

• Jointly learning A/V type classification and vascular tree disentanglement significantly enhances A/V separation accuracy.
• The proposed multi-task Siamese network with CAV provides a more robust and accurate solution for retinal image analysis.
• This approach offers a promising direction for automated diagnosis and monitoring of retinal vascular conditions.