Sparse diffusion models for multi-annotator medical image segmentation

Medical image segmentation is crucial for accurate medical diagnosis and treatment planning, but it often suffers from ambiguity due to the inherent complexity of medical images and variability among annotators. Addressing this issue requires to capture and integrate the diverse insights from mul- tiple annotators. Denoising diffusion models show advantages in predicting the annotations from multiple experts. However, the sampling process of these models is time-consuming. In this paper, we introduce a novel Sparse Diffusion-based Segmentation framework (SDSeg) for efficient multi-rater medical image segmentation, aiming to generate consensus masks at different levels of agreement among annotators. In the proposed framework, we build a Background-Adaptive Spatial Sparse (BASS) module that accelerates the inference process by concentrating computational efforts on critical regions within the image. The proposed module reduces the sampling time while maintaining high segmentation accuracy. Experimental results demonstrate the effectiveness of our method on five different datasets, showing significant improvements in segmentation reliability and efficiency compared to existing techniques. Code will be released via https://github.com/lhaof/SparseDiffusion after the acceptance.