Microvasculature Segmentation and Intercapillary Area Quantification of the Deep Vascular Complex Using Transfer...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a microvascular complication of diabetes affecting retinal circulation.
• Optical coherence tomography angiography (OCT-A) is crucial for visualizing these retinal changes.

Purpose of the Study:

• To demonstrate accurate segmentation of superficial capillary plexus (SCP) and deep vascular complex (DVC) using a convolutional neural network (CNN).
• To enable quantitative analysis of retinal vascular morphology and perfusion in patients with DR.

Main Methods:

• Retinal OCT-A images (6x6 mm FOV) were used for CNN training.
• Transfer learning was employed, leveraging a CNN pre-trained on smaller FOVs.
• Automated segmentation of SCP and DVC was performed for quantitative perfusion analysis.

Main Results:

• The CNN achieved high accuracy and Dice indices for segmenting both SCP (0.8599, 0.8618) and DVC (0.7986, 0.8139).
• Automated segmentations exceeded inter-rater reliability for both SCP and DVC.
• The method successfully maintained distinct vascular morphologies.

Conclusions:

• Transfer learning enables high-quality automated segmentation with reduced manual annotation needs.
• The developed CNN provides a robust tool for quantitative analysis of retinal perfusion in DR.
• Accurate microvasculature segmentation enhances perfusion analysis for diabetic retinopathy.