ESRVS: Extreme Semi-Supervised Retinal Vessel Segmentation with a Single Annotated Image

Researchers proposed ESRVS, which propagates retinal vessel labels from just one manually annotated image plus unlabeled data.

The method combines adapted DINOv3 and Mask2Former features, prototype cues, pseudo-label training, and adversarial refinement.

On eight public datasets, ESRVS outperformed prior semi-supervised methods and retained most of the fully supervised performance.

The work suggests retinal vessel segmentation can approach state-of-the-art accuracy with dramatically less annotation effort.