Balancing the Role of Priors in Multi-Observer Segmentation Evaluation

Area of Science:

• Medical Image Analysis
• Computational Pathology
• Computer-Aided Diagnosis

Background:

• Evaluating and combining multiple segmentations is complex.
• Existing methods like STAPLE rely heavily on truth priors and can ignore observer performance.
• Accurate truth prior modeling is challenging.

Purpose of the Study:

• To propose a Bayesian decision formulation for multi-observer segmentation evaluation.
• To develop a flexible framework integrating truth and observer priors.
• To address limitations of existing methods, particularly STAPLE.

Main Methods:

• Developed a Bayesian decision formulation for segmentation evaluation.
• Proposed a maximum a posteriori (MAP) principle for combining segmentations.
• Implemented a web-based software application for digitized uterine cervix images.
• Considered four scenarios: truth prior, observer prior, neither, or both.

Main Results:

• The proposed framework effectively integrates different priors for multi-observer segmentation.
• The method respects observer priors even without truth priors.
• Experimental results show favorable comparisons to STAPLE and Majority Vote Rule.
• The software application provides a flexible segmentation evaluation tool.

Conclusions:

• The Bayesian decision formulation offers a flexible and robust approach to multi-observer segmentation evaluation.
• The MAP-based method enhances segmentation accuracy by incorporating observer performance.
• The developed framework and software advance the field of medical image segmentation analysis.