Adaptive Annotation Correlation Based Multi-Annotation Learning for Calibrated Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Automated medical image segmentation is crucial for clinical applications.
• Current methods heavily depend on manual annotations, introducing subjectivity and bias.
• Existing approaches overlook the valuable information within multiple annotations.

Purpose of the Study:

• To propose a novel method, Adaptive annotation CorrelaTion based multI-annOtation LearNing (ACTION), for calibrated medical image segmentation.
• To address the limitations of existing methods by modeling correlations between multiple annotations.

Main Methods:

• ACTION employs consensus feature learning to exploit complementary information across annotations.
• Dynamic adaptive weighting is used to emphasize discriminative information within each annotation.
• Memory accumulation-replay integrates prior knowledge for multi-annotation settings.

Main Results:

• ACTION demonstrated superior performance on two medical image benchmarks across different modalities.
• The method effectively leverages complementary and discriminative information from multiple annotations.
• Experimental results show significant improvements over state-of-the-art methods.

Conclusions:

• The proposed ACTION method offers a robust solution for calibrated medical image segmentation.
• Modeling annotation correlations is key to overcoming subjectivity and bias in segmentation.
• ACTION provides a promising direction for developing more reliable automated medical image analysis tools.