Resdiff: Residual Diffusion Model for Conditional Distribution Learning in Medical Image Segmentation

Learning conditional distribution with multiple plausible hypotheses is significant to address the inherent ambiguity in medical image segmentation. Deep generative models have demonstrated their remarkable superiority in conditional distribution learning. However, existing generative-based approaches tend to generate hypotheses with limited diversities and are difficult to train, resulting in poorly calibrated segmentation results. Recently, diffusion-based generative models have exhibited outstanding performance in various tasks and have the potential to resolve these problems. In this paper, we introduce a novel residual diffusion model (ResDiff) that overcomes these problems and significantly improves model calibration. The proposed ResDiff first takes one-hot encoded segmentation label map as diffusion targets in continuous space, enabling compatibility with the existing diffusion model paradigm in segmentation tasks. Furthermore, we extend the vanilla diffusion models by introducing the residual learning strategy in the diffusion process, which dramatically improves sampling efficiency and model calibration. Moreover, we further improve the training process by adding deep supervisions in the intermediate steps of the diffusion model. We evaluate the ResDiff on the public BraTS2018 and INSTANCE2022 datasets. The experimental results demonstrate that the ResDiff outperforms existing conditional distribution learning methods and achieves state-of-the-art results.