Bayes-STAPLE: a python module for Bayesian label fusion

Abstract

Label fusion is widely used in fields such as medical imaging, remote sensing, and product rating, among others.
In a label-fusion application, usually several assessments of the same item are obtained, often from different algorithms or human experts, to estimate the missing ground truth.
However, a quantification of performance uncertainty is important to understand whether more data (e.g., more raters) are needed for a more certain estimation of the ground truth.
In this work, we describe a software that implements our Bayesian extension of STAPLE in order to provide an estimation of uncertainty, which could help decision makers in accepting estimations or requiring additional data.
The main feature of the developed software is its wider applicability in many contexts where ground truth is not available, but can be estimated by reaching a consensus among different rates.
The experimental results on the MSSEG2016 dataset on brain segmentations to identify multiple sclerosis lesions are reported to show the effectiveness of the methods and the use of the software.