Reproduce “Annotator Consensus Prediction for Medical Image Segmentation with Diffusion Models” on OpenResearch

About this paper

Tomer Amit, Shmuel Shichrur, Tal Shaharabany, Lior Wolf

Researchers from Tel-Aviv University developed a conditional diffusion model for medical image segmentation that explicitly models and averages different levels of annotator consensus to generate robust segmentation maps. This approach achieved state-of-the-art soft Dice scores across five QUBIQ benchmark datasets, consistently outperforming previous multi-annotator prediction methods.

Key insights

• Explicitly modeling and training a diffusion model to predict individual consensus levels (e.g., at least 'c' annotators agree) and then averaging these predictions at inference is more effective than directly predicting a soft-label map or aggregating individual annotator outputs.
• Employing multiple generations during the stochastic denoising inference process significantly stabilizes the output and improves segmentation performance by averaging out inherent variability.
• The method's performance positively correlates with the mean pairwise annotator agreement within a dataset, demonstrating its ability to leverage higher expert consensus effectively.

Summary by alphaXiv.