<p>Foundation models are pretrained on massive datasets. However, collecting medical datasets is expensive and time-consuming, and raises privacy concerns. Here we show that synthetic data generated via conditioning with disease labels can be leveraged for building high-performing medical foundation models. We pretrained a retinal foundation model, first with approximately one million synthetic retinal images with physiological structures and feature distribution consistent with real counterparts, and then with only 16.7% of the 904,170 real-world colour fundus photography images required in a recently reported retinal foundation model (RETFound). The data-efficient model performed as well or better than RETFound across nine public datasets and four diagnostic tasks; and for diabetic-retinopathy grading, it used only 40% of the expert-annotated training data used by RETFound. We also support the generalizability of the data-efficient strategy by building a classifier for the detection of tuberculosis on chest X-ray images. The text-conditioned generation of synthetic data may enhance the performance and generalization of medical foundation models.</p>