A Semi-Supervised Retinal Vessel Segmentation Method via Adaptive Uncertainty Estimation

We introduce a semi-supervised vessel segmentation technique that leverages a minimal amount of labeled data alongside substantial unlabeled data. This method addresses the limitations of supervised learning in medical image segmentation, which typically requires labor-intensive pixel-level labeling by experts. Using semi-supervised learning, our proposed adaptive uncertainty estimation (AUE) method enhances model performance through pixel-level uncertainty estimation and adaptive thresholding. This technique improves predictive accuracy by preserving high-confidence pixels between teacher-student networks, thereby effectively utilizing unlabeled data to acquire new features. Our approach surpasses both supervised and other semi-supervised models in accuracy on the STARE public retinal dataset. We have also benchmarked against several advanced semi-supervised segmentation methods, with our method achieving the best performance.