Contrastive self-supervised learning from 100 million medical images with optional supervision.

Area of Science:

• Artificial Intelligence in Medicine
• Medical Image Analysis
• Machine Learning

Background:

• Accurate artificial intelligence (AI) for medical image assessment requires extensive annotated datasets.
• Annotation is costly and time-consuming, often needing expert radiologists.
• Existing methods struggle with the scale and complexity of medical imaging data.

Purpose of the Study:

• To develop a self-supervised learning method for training AI on large-scale medical image datasets.
• To overcome the limitations of costly manual annotation in medical AI development.
• To enable learning from over 100 million medical images across various modalities.

Main Methods:

• Utilized contrastive learning and online feature clustering.
• Trained on a massive dataset of over 100 million medical images (radiography, CT, MR, US).
• Applied learned features to guide supervised and hybrid self-supervised/supervised model training for downstream tasks.

Main Results:

• Achieved significant accuracy increases (3-7% AUC boost) in detecting abnormalities in chest radiography and brain CT hemorrhage.
• Accelerated model convergence by up to 85% compared to training without pretraining.
• Demonstrated increased robustness to image augmentations and variations.

Conclusions:

• The proposed self-supervised approach significantly enhances accuracy and robustness in medical image assessment.
• Outperforms state-of-the-art methods trained on medical or general vision datasets (e.g., ImageNet).
• Enables efficient and effective AI model development for complex medical imaging challenges.