[Paper Review] VascX Models: Model Ensembles for Retinal Vascular Analysis from Color Fundus Images

Jose Vargas Quiros, Bart Liefers+5 more|arXiv (Cornell University)|Sep 24, 2024

Retinal Imaging and AnalysisMedicine3 citations

TL;DR

VascX models introduce a robust, multi-model ensemble framework for automated retinal vascular analysis in color fundus images (CFIs), leveraging diverse public and clinical datasets—including over 30 years of multi-vendor CFI data from the Rotterdam Study—to achieve state-of-the-art performance in vessel, artery-vein, optic disc, and fovea segmentation. The models outperform existing systems, especially on intermediate-quality images common in large cohort studies, and deliver higher precision than human graders in vessel segmentation.

ABSTRACT

We introduce VascX models, a comprehensive set of model ensembles for analyzing retinal vasculature from color fundus images (CFIs). Annotated CFIs were aggregated from public datasets . Additional CFIs, mainly from the population-based Rotterdam Study were annotated by graders for arteries and veins at pixel level, resulting in a dataset diverse in patient demographics and imaging conditions. VascX models demonstrated superior segmentation performance across datasets, image quality levels, and anatomic regions when compared to existing, publicly available models, likely due to the increased size and variety of our training set. Important improvements were observed in artery-vein and disc segmentation performance, particularly in segmentations of these structures on CFIs of intermediate quality, common in large cohorts and clinical datasets. Importantly, these improvements translated into significantly more accurate vascular features when we compared features extracted from VascX segmentation masks with features extracted from segmentation masks generated by previous models. With VascX models we provide a robust, ready-to-use set of model ensembles and inference code aimed at simplifying the implementation and enhancing the quality of automated retinal vasculature analyses. The precise vessel parameters generated by the model can serve as starting points for the identification of disease patterns in and outside of the eye.

Motivation & Objective

• To develop a robust, generalizable model ensemble for retinal vascular analysis that performs consistently across diverse imaging devices and image quality levels.
• To address the limitations of existing models trained on small, homogeneous datasets by incorporating a large, diverse dataset from the population-based Rotterdam Study with multi-vendor, multi-year CFI acquisitions.
• To improve segmentation accuracy—particularly for artery-vein and optic disc structures—on intermediate-quality images, which are prevalent in large cohort and clinical datasets.
• To outperform existing automated systems like Automorph and LWNet in segmentation and derived vascular feature extraction.
• To provide a ready-to-use, open-source inference pipeline that simplifies deployment and enhances reproducibility in retinal image analysis research.

Proposed method

• The models were trained on a combination of public datasets and newly annotated CFIs from the Rotterdam Study, including pixel-level annotations for arteries, veins, optic disc, and fovea.
• A multi-model ensemble approach was employed, using five-fold cross-training to improve robustness and generalization across diverse imaging conditions.
• Deep learning architectures such as U-Net variants were used, combined with strong data augmentation strategies to enhance performance on low- and intermediate-quality images.
• The inference pipeline includes standardized preprocessing steps, such as contrast normalization and image resizing, to ensure consistency across external datasets.
• The system integrates separate models for vessel, artery-vein, optic disc, and fovea segmentation, enabling modular and scalable analysis.
• A comprehensive evaluation framework was applied, comparing model outputs against human grader annotations using metrics like Dice score, mean absolute error, and Pearson correlation for derived vascular features.

Experimental results

Research questions

• RQ1 Can a model ensemble trained on diverse, real-world CFI data from multiple devices and over 30 years of imaging outperform existing systems in retinal vascular segmentation?
• RQ2 How does model performance vary across different image quality levels, particularly on intermediate-quality images common in large cohort studies?
• RQ3 To what extent do improved segmentation results translate into more accurate extraction of vascular biomarkers such as caliber, tortuosity, and bifurcation counts?
• RQ4 Can the VascX models surpass human graders in vessel segmentation precision, especially in complex or low-contrast images?
• RQ5 How effective is the VascX pipeline in preserving vascular tree connectivity and structural integrity in artery-vein segmentation?

Key findings

• VascX models achieved superior segmentation performance across all metrics, with Dice scores of 0.834 for vessel segmentation, 0.772 for artery-vein segmentation, and 0.733 for optic disc segmentation on the Rotterdam Study dataset.
• The largest performance gains were observed in artery-vein segmentation on intermediate-quality images, where the model significantly outperformed prior systems.
• VascX outperformed human graders in vessel segmentation precision, with a mean absolute error of 0.005 mm for vessel caliber compared to grader annotations.
• The model demonstrated high correlation (Pearson r = 0.791) with ground truth for bifurcation counts in veins, indicating strong fidelity in capturing complex vascular features.
• Derived vascular features from VascX showed lower mean absolute error (0.005–0.006 mm) and higher correlation (r = 0.757–0.791) with grader-based features than Automorph and LWNet.
• The VascX inference pipeline, available via GitHub, enables easy deployment on external datasets with consistent preprocessing and model inference, enhancing reproducibility and adoption.