Revisiting the evaluation of segmentation results: introducing confidence maps.

We introduce a novel framework, called Confidence Maps Estimating True Segmentations (Comets), to store segmentation references for medical images, combine multiple references, and measure the discrepancy between a segmented object and a reference. The core feature is the use of efficiently encoded confidence maps, which reflect the local variations of blur and the presence of nearby objects. Local confidence values are defined from expert user input, and used to define a new discrepancy error measure, aimed to be directly interpreted quantitatively and qualitatively. We illustrate the use of this framework to compare different segmentation methods and tune a method's parameters.