A Soft Label Method for Medical Image Segmentation with Multirater Annotations.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Multirater annotations are crucial for mitigating diagnostic errors in medical image analysis.
• Soft labels, derived from averaging rater annotations, are commonly used but their full potential is underexplored.
• Existing methods often struggle with regions of high uncertainty and image complexity.

Purpose of the Study:

• To improve soft label-based medical image segmentation by leveraging interpretable information from multiple raters.
• To reduce reliance on uncertain local soft labels and enhance focus on image features.
• To address challenges in regions with weak annotation supervision and high image difficulty.

Main Methods:

• Introduced local self-ensembling learning with consistency regularization to prioritize image features.
• Utilized pixelwise interclass variance to identify regions of high uncertainty.
• Applied label smoothing to annotations to mitigate overconfidence in structural edges.

Main Results:

• Achieved a 4.2% accuracy improvement on a synthetic dataset and 2.7% on a fundus dataset compared to the soft label baseline.
• Demonstrated consistent outperformance against existing multirater strategies and state-of-the-art methods.
• Showcased improvements without introducing additional model parameters.

Conclusions:

• The proposed method offers a simple yet effective solution for medical image segmentation with multirater annotations.
• Leveraging interpretable information from soft labels can significantly enhance segmentation performance.
• The approach effectively handles uncertainty and complexity in clinical image analysis.