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 augmentation methods often fail to capture realistic domain variations in medical imaging, such as scanner differences.

Purpose of the Study:

• Introduce counterfactual contrastive learning, a novel framework using causal image synthesis for improved positive pair generation.
• Enhance robustness to acquisition shifts and improve downstream performance in medical image analysis.

Main Methods:

• Leveraged causal image synthesis to generate counterfactual positive pairs that capture relevant domain variations.
• Evaluated the framework across five datasets (chest radiography, mammography) using SimCLR and DINO-v2 objectives.
• Assessed performance based on robustness to acquisition shifts and downstream task accuracy.

Main Results:

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

Conclusions:

• Counterfactual contrastive learning offers a powerful approach to generating semantically meaningful positive pairs for medical imaging.
• The method significantly improves model generalization, robustness, and fairness in medical image analysis tasks.
• This framework holds promise for developing more reliable and equitable AI in healthcare.