VascX Models: Model Ensembles for Retinal Vascular Analysis from Color Fundus Images

We introduce VascX models, a comprehensive set of model ensembles foranalyzing retinal vasculature from color fundus images (CFIs). Annotated CFIswere aggregated from public datasets . Additional CFIs, mainly from thepopulation-based Rotterdam Study were annotated by graders for arteries andveins at pixel level, resulting in a dataset diverse in patient demographicsand imaging conditions. VascX models demonstrated superior segmentationperformance across datasets, image quality levels, and anatomic regions whencompared to existing, publicly available models, likely due to the increasedsize and variety of our training set. Important improvements were observed inartery-vein and disc segmentation performance, particularly in segmentations ofthese structures on CFIs of intermediate quality, common in large cohorts andclinical datasets. Importantly, these improvements translated intosignificantly more accurate vascular features when we compared featuresextracted from VascX segmentation masks with features extracted fromsegmentation masks generated by previous models. With VascX models we provide arobust, ready-to-use set of model ensembles and inference code aimed atsimplifying the implementation and enhancing the quality of automated retinalvasculature analyses. The precise vessel parameters generated by the model canserve as starting points for the identification of disease patterns in andoutside of the eye.