Related Experiment Video
Updated: Jan 8, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.3K
Meta-learners for few-shot weakly-supervised optic disc and cup segmentation on fundus images
Pandega Abyan Zumarsyah1, Igi Ardiyanto1, Hanung Adi Nugroho1
1Department of Electrical and Information Engineering, Faculty of Engineering, Universitas Gadjah Mada, Grafika St. No. 2, Yogyakarta, 55281, Indonesia.
Computers in Biology and Medicine
|December 18, 2025
Summary
This study introduces Efficient Omni ProtoSeg (EO-ProtoSeg) for glaucoma diagnosis, achieving high accuracy in optic disc and optic cup segmentation with minimal sparse labels. This few-shot weakly-supervised segmentation approach offers a lightweight and effective solution for low-data scenarios.
Area of Science:
- Medical Imaging
- Computer Vision
- Machine Learning
Background:
- Accurate optic disc (OD) and optic cup (OC) segmentation is critical for glaucoma diagnosis.
- Limited labeled data presents a significant challenge for developing robust segmentation models.
Purpose of the Study:
- To develop advanced meta-learners for few-shot weakly-supervised segmentation (FWS) to address data scarcity in OD/OC segmentation.
- To improve existing meta-learning frameworks by enhancing data utilization and computational efficiency.
Main Methods:
- Introduced Omni meta-training to improve data utilization and shot diversification in meta-learning.
- Developed efficient meta-learner versions to reduce computational costs without compromising performance.
- Implemented sparsification techniques to simulate various types of sparse labels (scribbles, points, regions).
Main Results:
- The proposed Efficient Omni ProtoSeg (EO-ProtoSeg) significantly outperformed existing meta-learners on DRISHTI-GS, REFUGE, and RIM-ONE r3 datasets.
- EO-ProtoSeg achieved high intersection over union (IoU) scores (88.15% OD, 71.17% OC on REFUGE) with only one sparsely labeled image.
- Achieved performance comparable to unsupervised domain adaptation methods with significantly fewer parameters and no retraining requirement.
Conclusions:
- Few-shot weakly-supervised segmentation (FWS) is a promising, lightweight, and effective approach for medical image segmentation in low-data regimes.
- EO-ProtoSeg demonstrates the potential of advanced meta-learning strategies for efficient and accurate glaucoma diagnosis support.
