Revisiting the evaluation of segmentation results: introducing confidence maps

Area of Science:

• Medical image analysis
• Computer vision
• Biomedical engineering

Background:

• Accurate medical image segmentation is crucial for diagnosis and treatment planning.
• Evaluating segmentation accuracy often relies on subjective or limited reference comparisons.

Purpose of the Study:

• Introduce Confidence Maps Estimating True Segmentations (Comets) for robust medical image segmentation reference management.
• Develop a quantitative measure of segmentation discrepancy using confidence maps.
• Facilitate comparison and parameter tuning of different segmentation methods.

Main Methods:

• Utilized efficiently encoded confidence maps reflecting local image variations (blur, object proximity).
• Incorporated expert user input to define local confidence values.
• Developed a novel discrepancy error measure interpretable both quantitatively and qualitatively.

Main Results:

• Demonstrated the framework's ability to store and combine multiple segmentation references.
• Showcased the utility of confidence maps for assessing segmentation accuracy.
• Successfully applied the framework to compare and tune segmentation methods.

Conclusions:

• Comets provides a novel and interpretable framework for medical image segmentation evaluation.
• Confidence maps offer a powerful tool for quantifying segmentation quality.
• The framework enhances the development and validation of medical image segmentation algorithms.