Multi-rater Prism: Learning self-calibrated medical image segmentation from multiple raters.

Area of Science:

• Medical image analysis
• Computer vision
• Machine learning

Background:

• Medical image segmentation often requires multiple expert annotations to reduce bias.
• Standard deep learning models struggle with multi-annotator datasets.
• Addressing inter-observer variability is crucial for robust segmentation.

Purpose of the Study:

• To propose a novel neural network framework, Multi-rater Prism (MrPrism), for medical image segmentation using multiple expert labels.
• To develop a method that learns from inter-observer variability and image semantics simultaneously.
• To achieve self-calibrated segmentation results that reflect expert agreement.

Main Methods:

• Introduced Multi-rater Prism (MrPrism), a recurrent neural network framework inspired by iterative half-quadratic optimization.
• Developed Converging Prism (ConP) for calibrated segmentation and Diverging Prism (DivP) for multi-rater confidence map estimation.
• Implemented an iterative process where ConP and DivP mutually enhance each other.

Main Results:

• The recurrent process of ConP and DivP demonstrated mutual improvement between segmentation and confidence estimation.
• MrPrism achieved superior performance compared to state-of-the-art methods across various medical image segmentation tasks.
• The framework successfully learned inter-observer variability and produced self-calibrated segmentation.

Conclusions:

• MrPrism offers an effective solution for medical image segmentation with multiple expert annotations.
• The proposed recurrent framework successfully models inter-observer variability and improves segmentation accuracy.
• This approach advances the field by enabling deep learning models to leverage consensus and disagreement in multi-rater data.