Statistical Fusion of Surface Labels Provided by Multiple Raters

Area of Science:

• Medical imaging analysis
• Computational anatomy
• Biomedical engineering

Background:

• Accurate delineation of anatomical structures is crucial for size and morphology studies.
• Reproducibility measures, often from repeated scans or delineations, are vital.
• Existing methods for estimating true structure and rater performance are primarily demonstrated on volumetric images.

Purpose of the Study:

• To extend simultaneous estimation methods for true structure and rater performance to 2D surfaces parameterized as triangle meshes.
• To address challenges in surface parameterization for label homogeneity using Markov random fields.
• To evaluate the method's effectiveness using simulated raters and real surface data.

Main Methods:

• Application of simultaneous estimation methods to 2D surfaces represented as triangle meshes.
• Enforcement of label homogeneity via a Markov random field with a tailored energy function.
• Utilizing simulated raters (global and boundary error models) and atlas-registered surface labels for validation.

Main Results:

• The developed method successfully estimates true segmentation on both simulated and real 2D surface data.
• The study analyzes the impact of rater accuracy using different error models.
• The Markov random field approach effectively handles surface parameterization challenges for label homogeneity.

Conclusions:

• The proposed method provides a robust framework for estimating accurate segmentations on 2D parameterized surfaces.
• This advancement enhances the reliability of reproducibility and variability measures in anatomical studies.
• The technique is valuable for analyzing anatomical structure morphology and size from surface-based data.