One-Vote Veto: Semi-Supervised Learning for Low-Shot Glaucoma Diagnosis.

Area of Science:

• Ophthalmology
• Computer Vision
• Medical Imaging

Background:

• Automated glaucoma diagnosis using Convolutional Neural Networks (CNNs) shows promise but requires extensive labeled data.
• Limited labeled data is a significant challenge in biomedical image classification, especially for rare diseases and expert-intensive labeling.

Purpose of the Study:

• To address data limitations in automated glaucoma diagnosis.
• To develop a low-shot learning method for training CNNs with limited and imbalanced datasets.
• To introduce a semi-supervised learning strategy to improve accuracy using unlabeled data.

Main Methods:

• Extended conventional Siamese networks to create a multi-task Siamese network (MTSN) capable of using various backbone CNNs.
• Introduced One-Vote Veto (OVV) self-training, a semi-supervised strategy specifically for MTSNs, leveraging both self-predictions and contrastive predictions on unlabeled data.
• Validated methods on a large dataset of 66,715 fundus photographs and three smaller clinical datasets.

Main Results:

• MTSN with limited training data achieved accuracy comparable to backbone CNNs trained on datasets 50 times larger.
• OVV self-training effectively utilized unlabeled data to fine-tune pre-trained MTSNs, enhancing diagnostic accuracy.
• The proposed methods demonstrated effectiveness and generalizability across diverse fundus image datasets.

Conclusions:

• The MTSN combined with OVV self-training offers a robust solution for low-shot, imbalanced glaucoma diagnosis from fundus images.
• These advancements significantly reduce the reliance on large labeled datasets, making automated diagnosis more accessible.
• The developed techniques show potential for broader applications in medical image analysis where data scarcity is an issue.