CaliDiff: Multi-rater annotation calibrating diffusion probabilistic model towards medical image segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Medical image segmentation is vital for diagnostics and treatment planning.
• Current methods struggle to fully integrate diverse expert insights and can introduce bias.
• Deep learning models often aggregate expert labels, embedding inherent biases.

Purpose of the Study:

• To introduce CaliDiff, a novel multi-rater annotation calibration diffusion probabilistic model.
• To leverage diverse expert knowledge and refine annotations for improved medical image segmentation.
• To enhance the reliability and objectivity of medical diagnostics through advanced annotation calibration.

Main Methods:

• CaliDiff employs a multi-stage process involving shared-parameter inverse diffusion for bias normalization.
• Expertness Consistent Alignment minimizes annotation variance and enhances consistency in high-confidence regions.
• Committee-based Endogenous Knowledge Learning uses adversarial soft supervision for pseudo-ground truth generation, integrating Cross-Expert Fusion and Implicit Consensus Inference.

Main Results:

• CaliDiff significantly improves the calibration of multi-rater annotations in medical images.
• The model achieves state-of-the-art performance in medical image segmentation tasks.
• Experimental evaluations demonstrate enhanced reliability and objectivity in diagnostic outcomes.

Conclusions:

• CaliDiff effectively addresses limitations in traditional multi-rater annotation and deep learning aggregation methods.
• The proposed model enhances the utilization of expert knowledge for more accurate medical image segmentation.
• CaliDiff represents a significant advancement in improving the quality and trustworthiness of AI-driven medical diagnostics.