Diffusion models for medical imaging: the path to fairer, more robust and private models (52:10)

In this talk, we showed that advances in generative models can help mitigate this unmet need in a steerable fashion, algorithmically enriching our training dataset with synthetic examples that address shortfalls of underrepresented conditions or subgroups. We will show that generative models can automatically learn realistic augmentations from data in a label-efficient manner. We will further discuss how differential privacy can be used to protect diffusion models from risks such as membership attacks, while still producing synthetic images that are useful for downstream tasks.