Taking a Deep Look at Multi-rater Agreement for Calibrated Medical Image Segmentation
IEEE Transactions on Medical Imaging
|July 17, 2026
Summary
This study introduces MRNet+, a novel method for medical image segmentation that models multi-rater disagreement and expert variability. MRNet+ enhances segmentation accuracy by leveraging expert knowledge and inter-rater agreement patterns, outperforming existing methods.
Area of Science:
- Medical Image Analysis
- Computer Vision
- Artificial Intelligence
Background:
- Multi-rater annotations are crucial for accurate medical image analysis but often simplified in research.
- Conventional methods overlook valuable information in expert agreement and disagreement patterns.
Purpose of the Study:
- To develop a novel method, MRNet+, that explicitly models multi-rater (dis-)agreement and varying expertise levels in medical image segmentation.
- To improve segmentation performance by incorporating rich information from multiple expert annotations.
Main Methods:
- MRNet+ incorporates an Expertise-aware Inferring Module (EIM) to utilize rater expertise.
- A disentangled erasing-based Multi-rater Knowledge Modeling (MKM) module infers interobserver variability.
- A Multi-rater Perception Module (MPM) leverages disagreement cues to enhance segmentation.
Main Results:
- MRNet+ is the first method to generate calibrated predictions accounting for rater expertise in medical image segmentation.
- Achieves real-time performance (89 FPS on 256x256 images).
- Consistently outperforms state-of-the-art methods across ten diverse medical segmentation tasks.
Conclusions:
- MRNet+ effectively models multi-rater disagreement and expertise, significantly improving medical image segmentation.
- The framework demonstrates broad applicability and superior performance in real-world scenarios.
