Few-shot Weighted Style Matching for Glaucoma Detection
- Publisher:
- Springer
- Publication Type:
- Conference Proceeding
- Citation:
- Artificial Intelligence, 2022, 13069 LNAI, pp. 289-300
- Issue Date:
- 2022-01-01
Closed Access
Copyright Clearance Process
- Recently Added
- In Progress
- Closed Access
This item is closed access and not available.
Glaucoma is a harmful eye disease that can lead to irreversible blindness. Color fundus photography (CFP) is the most popular non-invasive and low-cost imaging modality to detect glaucoma. However, diagnosing glaucoma from fundus images is not an easy task and only clinicians with years of experience can do it. Deep neural network (DNN), especially convolutional neural network (CNN), has shown great power in medical image processing and has great potential for efficient glaucoma diagnosis. Nevertheless, fundus images captured by different cameras and devices may have different characteristics, which causes domain shift problem and severely affects CNN generalizing across different datasets. In this paper, we exploit unsupervised domain adaptation to address domain shift. Here, we assume only few target unlabeled samples are available, which is a more realistic and challenging problem. We present a novel few-shot weighted style matching framework (few-shot WSM) to robustly detect glaucoma from different fundus image datasets. The few-short WSM module reduces domain shift by matching the style of the source domain images and the target domain images. Experiment results show our framework effectively reduces the domain shift problem and significantly improves the glaucoma classification performance.
Please use this identifier to cite or link to this item: