Robust and data-efficient generalization of self-supervised machine learning for diagnostic imaging.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• Machine learning models in medicine can perform comparably to human experts.
• Model performance often degrades significantly when applied to data outside their training distribution.
• This limits the real-world applicability of AI in healthcare.

Purpose of the Study:

• To introduce a novel representation-learning strategy, REMEDIS, for medical imaging tasks.
• To enhance model robustness and training efficiency, particularly in out-of-distribution scenarios.
• To mitigate performance degradation in machine learning models when encountering new data.

Main Methods:

• REMEDIS combines large-scale supervised transfer learning on natural images.
• It incorporates intermediate contrastive self-supervised learning on medical images.
• The strategy requires minimal task-specific customization.

Main Results:

• REMEDIS improved in-distribution diagnostic accuracies by up to 11.5% compared to supervised baselines.
• In out-of-distribution settings, REMEDIS required only 1-33% of data for retraining to match baseline performance.
• The strategy demonstrated utility across six imaging domains and 15 test datasets.

Conclusions:

• REMEDIS offers a robust and efficient approach to machine learning in medical imaging.
• The strategy effectively addresses the challenge of out-of-distribution performance degradation.
• REMEDIS has the potential to accelerate the development and deployment of AI tools in medical diagnostics.