Abstract—Segmentation of the Optic Disc (OD) and Optic Cup (OC) on fundus images is vital for glaucoma diagnosis. Deep learning has been utilized to automate this segmentation task. However, it typically requires a large number of labeled images. To address this issue, we propose Dual Metric Learning for OD and OC Segmentation (DMLOS), which requires only a few images with some of their pixels labeled. It consists of a neural network for embedding extraction, followed by dual branches to obtain prototypes and predictions. The Omni Training algorithm is used to improve data utilization and use a diverse number of shots. Meanwhile, DeepLabv3+ and miniUNet are the neural network used. We extensively evaluated DMLOS on the DRISHTI-GS, REFUGE, and RIM-ONE r3 datasets using various numbers of shots and label densities. Using 15 shots with 0.1 grid density, DMLOS achieved an Intersection over Union of 92.56% for OD and 73.08% for OC on DRISHTI-GS, surpassing other less-supervised methods. The results demonstrate the potential of DMLOS as an effective approach in low-label scenarios.
 
Keywords—few-shot, weakly-supervised, prototypes, sparse label, fundus image, segmentation, glaucoma


Copyright © 2026 by the authors. This is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited (CC BY 4.0).