Automated glaucoma assessment often relies on segmenting the optic disc (OD) and optic cup (OC) from fundus images. However, models trained on limited labeled datasets tend to perform poorly when applied to images from new domains, a common challenge in medical imaging. A new preprint on arXiv tackles this by combining semi-supervised teacher-student learning with a weighting scheme for pseudo-labels.
The paper's approach, as described in the abstract, centers on "pseudo-label foreground quality weighting." In teacher-student frameworks, the teacher generates pseudo-labels for unlabeled data, which the student then learns from. The proposed method weights these pseudo-labels based on the quality of their foreground regions, aiming to reduce the impact of noisy or unreliable labels during training.
The abstract is truncated, so the available text does not include experimental details or quantitative results. The contribution, based on the title and the visible portion, is the specific weighting mechanism for improving domain generalization in OD/OC segmentation. As this is a single preprint, the claims have not yet been peer-reviewed or independently verified.