Synthesizing Ultrasound Datasets using Mask Conditional Diffusion Models

Development of artificial intelligence (AI) enabled clinical applications requires large amount of labeled data. The robustness of these methods strongly hinges on the ability to cover a wide variety of patient subgroups, and image variations that are typically seen in real world. To alleviate the need to acquire and annotate large amounts of data to train AI models, this work demonstrates the utility of mask conditional diffusion models to synthesize ultrasound images. Specifically, we propose to generate ultrasound images of cardiac and thyroid anatomies.