A data-efficient strategy for building high-performing medical foundation models.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Collecting large medical datasets for foundation models is costly, time-consuming, and raises privacy issues.
• Existing medical foundation models require extensive real-world data for pretraining.

Purpose of the Study:

• To investigate the efficacy of synthetic data in training high-performing medical foundation models.
• To develop a data-efficient approach for pretraining medical foundation models.

Main Methods:

• Pretrained a retinal foundation model using approximately one million synthetic retinal images.
• Compared the synthetic data-trained model with a model trained on real-world data (RETFound), using only 16.7% of the real images.
• Evaluated model performance across nine public datasets and four diagnostic tasks.
• Validated the approach by building a tuberculosis detection classifier on chest X-ray images.

Main Results:

• The data-efficient model achieved performance comparable to or better than RETFound across multiple datasets and tasks.
• For diabetic retinopathy grading, the model utilized only 40% of the expert-annotated data compared to RETFound.
• Demonstrated generalizability by successfully training a tuberculosis detection classifier.

Conclusions:

• Synthetic data generated via disease label conditioning can effectively train high-performing medical foundation models.
• This data-efficient strategy enhances model performance and generalization while mitigating data acquisition challenges.
• Text-conditioned synthetic data generation shows promise for advancing medical AI.