Robust image representations with counterfactual contrastive learning

Area of Science:

• Medical Imaging
• Machine Learning
• Computer Vision

Background:

• Contrastive pretraining enhances model generalization but relies heavily on data augmentation for positive pairs.
• Existing methods use generic transformations that fail to capture realistic medical imaging domain variations like scanner differences.

Purpose of the Study:

• Introduce counterfactual contrastive learning, a novel framework using causal image synthesis for generating realistic positive pairs.
• Improve robustness to acquisition shifts and enhance downstream performance in medical imaging tasks.

Main Methods:

• Leveraged causal image synthesis to create counterfactual positive pairs that mimic relevant domain variations.
• Evaluated the framework on five datasets (chest radiography, mammography) with SimCLR and DINO-v2 objectives.
• Assessed performance against standard contrastive learning in terms of robustness and downstream task accuracy.

Main Results:

• Counterfactual contrastive learning demonstrated superior robustness to acquisition shifts compared to standard methods.
• Achieved better downstream performance on both in-distribution and external datasets, particularly for under-represented scanners.
• Showed reduction in subgroup disparities across biological sex, indicating broader applicability beyond acquisition shifts.

Conclusions:

• Counterfactual contrastive learning offers a more effective approach to data augmentation in medical imaging.
• The framework enhances model generalization, robustness, and fairness, particularly in diverse acquisition scenarios.
• This method holds promise for improving AI in healthcare by addressing domain shift and subgroup disparities.