Robust Bayesian fusion of continuous segmentation maps.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Analyzing multiple image labels or probability maps requires fusing data from various segmentation algorithms or human raters.
• Accurate fusion necessitates correctly weighting map combinations to account for rater agreement, outliers, and spatial uncertainty.

Purpose of the Study:

• To address shortcomings in prior continuous label fusion methods.
• To introduce a novel approach for jointly estimating a reliable consensus map and assessing rater reliability (outliers and confidence).

Main Methods:

• Utilizing heavy-tailed distributions (Laplace, Student's t, generalized double Pareto) for robust estimation of local rater performance, outperforming classical Gaussian likelihood.
• Employing variational calculus and scale mixture representations for a unified, tractable inference scheme.
• Incorporating bias and spatial priors for accurate rater bias estimation and consensus map smoothness control.

Main Results:

• Demonstrated a robust approach for estimating consensus maps and rater confidence.
• Successfully identified outliers and assessed individual rater performance.
• Showcased the ability to cluster raters using variational boosting, generating alternative consensus maps.

Conclusions:

• The proposed method offers a significant advancement in continuous label fusion for medical image analysis.
• The use of heavy-tailed distributions provides a more reliable way to handle disagreements and uncertainties in multi-rater segmentation.
• Validated effectiveness on MR prostate delineations and LIDC-IDRI lung nodule segmentations.