Deep learning models for retinal vascular analysis from color fundus images

Abstract

Purpose : The retinal vascular network is known to be an indicator of vascular health which can be imaged through non-invasive, increasingly affordable means such as fundus photography. This calls for automated vascular feature extraction pipelines robust to differences in image quality. We developed and evaluated models for vessel, artery-vein and optic disc segmentation from color fundus images (CFI) of varying quality. These models are the foundation of our vascular analysis pipeline, which relies on segmentation masks (Fig 1).

Methods : We collected a development set of 329 CFIs; with 85 CFIs from publicly available datasets HRF (45) and AV-DRIVE (40) and 244 CFIs (F1 and F2) sampled from the Rotterdam Study (RS) dataset. We only discarded CFIs considered unusable by the graders to capture a variety of image conditions. RS CFIs were annotated by four graders using a custom interface initialized with output from a vessel segmentation model. All CFIs were annotated for artery-vein crossings. We trained custom U-Net models on this set (5-fold cross-validation), split into 70% training, 10% validation and 20% test. Foreground Dice scores were calculated for evaluation. We compared our models with the state-of-the-art models in the Automorph package for CFI analysis. We measured performance on RS CFIs only since Automorph was trained on the public sets. The optic disc model was trained on 2800 publically available CFI annotations.

Results : The vessel segmentation model achieved a Dice score of 0.855 ± 0.023 for the foreground, outperforming Automorph’s model (0.761). The artery-vein segmentation algorithm obtained AUCs of 0.765 ± 0.033 (arteries); 0.804 ± 0.027 (veins) and 0.532 ± 0.038 compared to 0.715 (arteries), 0.756 (veins) and 0.410 (crossings) for Automorph. The artery-vein model was visibly better able to recover connectivity of the vessel trees and the disc model was more consistent (Fig 2).

Conclusions : Our segmentation models improved over previous models when evaluated on a dataset of CFIs of varying quality. They are an important step towards more robust vascular analysis. The effect of CFI quality on end features may be further reduced with better CFI selection algorithms targeting vessel visibility.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

 

Stages of our analysis pipeline (actual model output).

 

From left to right: Automorph outputs, our models’ outputs, and separate artery and vein masks (our model).