VascX Models: Deep Ensembles for Retinal Vascular Analysis From Color Fundus Images.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate segmentation of retinal vasculature is crucial for diagnosing ophthalmic and systemic diseases.
• Existing deep learning models for retinal image analysis have limitations in performance and consistency.

Purpose of the Study:

• To present and validate VascX, a deep learning model ensemble for segmenting vessels, arteries-veins, optic discs, and localizing the fovea in color fundus images (CFIs).
• To make VascX preprocessing code and model weights publicly available to advance research on retinal vasculature.

Main Methods:

• Trained UNet model ensembles using a diverse dataset combining over 15 annotated datasets and CFIs from Dutch studies.
• Employed a robust preprocessing algorithm and strong data augmentations for model training.
• Compared VascX segmentation performance (Dice scores) and feature extraction agreement (MAE, Pearson correlation) against AutoMorph and LittleWNet models.

Main Results:

• VascX demonstrated superior performance across most datasets, particularly for artery-vein and optic disc segmentation.
• VascX exhibited consistent performance even with decreased image quality and in disc/fovea-centered images.
• VascX significantly improved the quality of vascular features, showing higher correlations with ground-truth features compared to existing models.

Conclusions:

• VascX models perform robustly across various conditions due to diverse training data.
• VascX offers improved segmentation quality, leading to more reliable vascular features for retinal vasculature analysis.
• Public availability of VascX aims to accelerate research linking retinal vascular biomarkers to health conditions.