Edge-Guided Contrastive Adaptation Network for Arteriovenous Nicking Classification Using Synthetic Data.

Area of Science:

• Ophthalmology
• Medical Imaging
• Cardiovascular Disease Research

Background:

• Retinal arteriovenous nicking (AVN) indicates systemic diseases, notably cardiovascular issues, and is associated with increased stroke risk.
• Current AVN classification is hindered by limited data and challenges in distinguishing subtle vein changes from normal variations.

Purpose of the Study:

• To develop an advanced method for accurate retinal arteriovenous nicking (AVN) classification.
• To address data scarcity and the subtle differences that challenge AVN detection.

Main Methods:

• A data synthesis technique was employed to generate diverse arteriovenous (AV) crossing images, including normal and AVN examples.
• An edge-guided unsupervised domain adaptation network was designed to bridge the gap between synthetic and real-world data.
• A semantic contrastive learning branch (SCLB) was introduced to focus on subtle venular width differences, using semantic triplets for training.

Main Results:

• The proposed method effectively mitigated the lack of data and domain shift issues between synthetic and real datasets.
• The semantic contrastive learning approach successfully addressed significant intra-class variations and minute inter-class differences in AVN.
• Experimental results demonstrated the outstanding performance of the developed method in AVN classification.

Conclusions:

• The novel approach enhances the accuracy of retinal arteriovenous nicking (AVN) detection by overcoming key data and classification challenges.
• This method holds potential for improved early identification of cardiovascular risks through retinal imaging analysis.