Leveraging Synthetic Medical Images for Effective Vision-Language Pre-training: Overcoming the Reliance on Real Image-Text Datasets
Synthetic medical images generated from real radiology reports can effectively substitute real images in vision-language pre-training, achieving comparable or even superior performance on downstream medical vision tasks.