Worst-Case Local Boundary Precision in Global Measures of Segmentation Reproducibility

The commonly used measures for reproducibility of semiautomatic/interactive image segmentation algorithms provide global estimates of the precision of the location of an object boundary in a group of segmentations. The joint Dice similarity coefficient, joint Tanimoto coefficient, generalized Tanimoto coefficient, coefficient of variation of volume, and intra-class correlation coefficient of volume are interpreted with respect to a new explicit measure of worst-case local object boundary precision. Experiments established 95% confidence intervals on this new measure for ranges of global reproducibility measures allowing global measures to be interpreted in terms of worst-case local precision. Joint Tanimoto coefficient and joint Dice coefficient are shown to be highly unstable over variations in the number of segmentations being compared. All of the existing measures of segmentation reproducibility are found to be flawed in a significant way with the exception of the generalized Tanimoto coefficient.