MedDataset: Synthesizing Infinite Image and Segmentation Data Using Diffusion Models

Deep learning models perform best when trained on large datasets, but in medical imaging, obtaining such data is challenging due to privacy concerns, scarcity of rare conditions, high costs, and time-intensive annotations. Generative models, particularly diffusion models, offer a promising solution by creating high-quality synthetic data with minimal effort. In this paper, we introduce MedDataset, an end-to-end framework that generates diverse synthetic medical images along with their segmentation masks without relying on any conditional inputs like masks or labels during image generation. Our approach leverages an unconditional diffusion model to produce images, and a novel Time-Gated Decoder that decodes the rich latent features into accurate segmentation masks by effectively utilizing information from different noise levels in the diffusion process. Remarkably, training the decoder requires only 10% of manually labeled images, yet achieves segmentation performance comparable to models trained on 100% labeled data. To our knowledge, this is the first work to propose an unconditional diffusion-based dataset generation model that jointly produces images and segmentation masks in an end-to-end manner. We conduct extensive qualitative and quantitative evaluations to demonstrate that our synthetic data closely matches real data. Additionally, experiments using U-Net and TransUNet on five different datasets show improved performance when training with both real and synthetic data from MedDataset. Code: https://github.com/Vibashan/MedDataset