Denoising Diffusion Medical Models

In this study, we introduce a generative model that can synthesizes a large number of radiographical image/label pairs and hence, is asymptotically favorable to downstream activities such as segmentation in bio-medical image analysis. Denoising Diffusion Medical Model (DDMM), the proposed technique, can create realistic X-ray images and associated segmentations on a small number of annotated datasets as well as other massive unlabeled datasets with no supervision. Radiograph/segmentation pairs are generated jointly by the DDMM sampling process in probabilistic mode.  As a result, a vanilla U-net that uses this data augmentation for segmentation task outperforms other kinds of similar data-centric approaches.