This paper discusses the use of denoising diffusion probabilistic models (DDPM) for medical image segmentation. The researchers highlight the challenges in this field, such as limited dataset sizes and the high variability of medical imaging data. They propose an improved ensembling scheme that uses smaller diffusion steps to create more robust latent representations. The model demonstrates better performance in domain-shifted settings while maintaining competitive in-domain performance. This study underscores the potential of DDPMs for semi-supervised medical image segmentation and offers insights into their performance optimization under domain shift.
Publication date: 14 Nov 2023
Project Page: arXiv:2311.07421v1
Paper: https://arxiv.org/pdf/2311.07421
