AV-casNet: Fully Automatic Arteriole-Venule Segmentation and Differentiation in OCT Angiography

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate segmentation and differentiation of retinal arteriole and venule (AV) are crucial for diagnosing eye diseases like diabetic retinopathy and systemic conditions such as hypertension.
• Optical coherence tomography angiography (OCTA) offers capillary-level blood flow data but lacks the distinct AV features seen in fundus photography.
• Current methods for AV differentiation in OCTA often require integration with other imaging modalities.

Purpose of the Study:

• To develop and evaluate a novel cascaded neural network for the automatic segmentation and differentiation of retinal AV solely from OCTA data.
• To overcome the limitations of OCTA in distinguishing between arterioles and venules without relying on additional imaging techniques.

Main Methods:

• A cascaded neural network architecture was proposed, comprising a convolutional neural network (CNN) for initial segmentation and a graph neural network (GNN) for refining connectivity.
• Various CNN and GNN architectures were explored and compared to determine optimal performance.
• The method was validated using multi-center clinical datasets encompassing both 3x3 mm² and 6x6 mm² OCTA scans.

Main Results:

• The proposed cascaded neural network demonstrated effective automatic segmentation and differentiation of retinal AV using OCTA.
• The GNN component successfully improved the connectivity of the initial segmentation generated by the CNN.
• Performance was evaluated across different OCTA scan sizes and datasets, indicating robustness.

Conclusions:

• The developed cascaded neural network provides a promising approach for automated AV segmentation and differentiation directly from OCTA.
• This technique has the potential to significantly enrich OCTA image information, aiding in the diagnosis of various ocular and systemic diseases.
• The method offers a standalone solution, reducing reliance on multimodal imaging for AV analysis.