On the Application of Human Rater Models to Statistical Fusion in Multi-Atlas Labeling
A. Asman, Antong Chen, and B. Landman. “On the Application of Human Rater Models to Statistical Fusion in Multi-Atlas Labeling.” In MICCAI 2011 Workshop on Multi-Atlas Methods and Statistical Fusion. Toronto, Canada, September 2011 (Oral Presentation) NIHMS317651
Full text: http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.701.3272&rep=rep1&type=pdf
Abstract
Segmentation is critical to understanding the complex
relationships between biological structure and function. Statistical algorithms to
fuse the output of multiple anatomical experts have been shown to be extremely
successful in improving combined accuracy and estimating individual rater
performance. In the interest of efficiency, human raters are commonly replaced
with label transfer via registration from pre-labeled atlas datasets (e.g., “multi-
atlas labeling”). Yet statistical methods based on models of human behavior are
often dramatically outperformed by simple voting techniques when applied to
multi-atlas fusion. Hitherto, the literature on fusion methods has focused on the
advantages of particular methods and has not presented clear exemplars to
illustrate the relative limitations. Herein, we perform a comparative
characterization of multiple statistical fusion algorithms based upon models of
human behavior and apply these characterizations to understanding multi-atlas
based segmentation accuracy.
