Metadata-enhanced contrastive learning from retinal optical coherence tomography images.

Area of Science:

• Medical imaging
• Deep learning
• Computer vision

Background:

• Deep learning models can automate disease screening and grading in medical images.
• Contrastive learning pretraining on natural images yields robust features for label-efficient medical image analysis.
• Conventional contrastive methods face challenges with medical image transformations and assumptions about image dissimilarity, especially in longitudinal datasets.

Purpose of the Study:

• To address limitations of conventional contrastive learning in medical imaging.
• To develop a metadata-enhanced strategy for improved feature extraction in medical image analysis.
• To leverage patient metadata to approximate true inter-image relationships for better contrastive pretraining.

Main Methods:

• Extended conventional contrastive frameworks with a metadata-enhanced strategy.
• Utilized patient identity, eye position, and time series data to approximate inter-image contrastive relationships.
• Applied the approach to large longitudinal datasets of retinal optical coherence tomography (OCT) images from patients with age-related macular degeneration (AMD).

Main Results:

• The metadata-enhanced approach outperformed standard contrastive methods and a retinal image foundation model in five out of six downstream tasks.
• Demonstrated benefits in both low-data and high-data regimes for AMD stage and type classification, and visual acuity prediction.
• Successfully incorporated temporal dynamics of disease progression into pretraining using metadata.

Conclusions:

• Metadata-enhanced contrastive learning offers a powerful strategy for medical image analysis, particularly for longitudinal studies.
• The method is modular, allowing for quick and cost-effective evaluation of metadata benefits in contrastive pretraining.
• This approach shows significant potential for advancing automated screening, monitoring, and grading of diseases like AMD.