Deep semi-supervised multiple instance learning with self-correction for DME classification from OCT images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Ophthalmology

Background:

• Diabetic macular edema (DME) recognition from optical coherence tomography (OCT) images is crucial for patient care.
• Supervised deep learning methods face challenges due to limited labeled OCT data for accurate DME analysis.
• Current approaches often treat DME identification as a multiple instance learning (MIL) problem, but do not leverage unlabeled data.

Purpose of the Study:

• To develop a novel deep semi-supervised multiple instance learning framework for DME recognition.
• To investigate the effectiveness of utilizing both limited labeled and extensive unlabeled OCT data.
• To reduce annotation costs while improving the accuracy of DME classification.

Main Methods:

• A deep semi-supervised multiple instance learning framework combining instance-level supervision with unlabeled data.
• A self-correction strategy using confidence-based pseudo-labeling and consistency regularization for noisy labels.
• A Student-Teacher architecture with consistency constraints to enhance model discrimination.

Main Results:

• The proposed framework significantly improves DME classification accuracy by incorporating unlabeled data.
• Outperforms existing multiple instance learning methods on large-scale DME OCT datasets.
• Demonstrates the feasibility of deep semi-supervised MIL for accurate DME detection at reduced annotation costs.

Conclusions:

• Deep semi-supervised multiple instance learning is a viable approach for DME recognition using OCT images.
• Leveraging unlabeled data alongside limited labeled data enhances classification performance.
• The proposed framework offers a cost-effective solution for accurate DME analysis.