Clinically Labeled Contrastive Learning for OCT Biomarker Classification.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Medical image analysis often requires extensive labeled data, which can be scarce.
• Clinical data is often more abundant than specialized biomarker labels.
• Optical coherence tomography (OCT) scans in ophthalmology show correlations between clinical values and biomarker structures.

Purpose of the Study:

• To develop a novel strategy for positive and negative set selection in contrastive learning for medical images.
• To leverage readily available clinical data as pseudo-labels for training AI models.
• To improve the direct classification of disease indicators from OCT scans using biomarker labels.

Main Methods:

• Utilized clinical data as pseudo-labels to select instances for supervised contrastive loss training.
• Trained a backbone network to align representation space with clinical data distribution.
• Fine-tuned the network using biomarker-labeled data with cross-entropy loss.
• Proposed a method using a linear combination of clinical contrastive losses.

Main Results:

• Achieved performance improvements of up to 5% in total biomarker detection AUROC.
• Demonstrated the effectiveness of using clinical data for pre-training in a novel setting with varying biomarker granularity.
• Successfully trained a network to learn representations aligned with clinical data distribution.

Conclusions:

• The proposed strategy effectively utilizes clinical data for contrastive learning in medical imaging.
• This method enhances the ability to classify disease indicators directly from OCT scans.
• The approach offers a promising direction for improving AI-driven diagnostics in ophthalmology with limited biomarker data.