Abstract
Validation of breast tissue segmentation algorithms is a vital step required to assess their suitability for the intended task such as biomechanical modeling of breast. A problem with such validation is that the true segmentation, a reference standard, is generally unavailable. Therefore, comparison of algorithmic segmentation with expert outlining of the breast tissues is preferably carried out for such validation. Often outlines from two or more experts are used to overcome the differences between different expert’s outlines of the same region i.e. inter-expert variability. However, the most appropriate approach with which to compare multiple expert outlines with algorithmic segmentation is so far unclear and several approaches are used in practice. We present a validation index (VI), a model based measure of the spatial overlap between algorithmic segmentation and multiple expert outlines. The levels of agreement i.e., the overlaps of experts’ outlines were used to define validation index (VI). The VI was evaluated using three idealized cases and data from a clinical study. The performance of VI with respect to three common formulations of the Dice similarity coefficient (DSC) was also studied. It was found that VI was more sensitive than the three DSCs to the change in agreement between experts and in all cases VI was smaller than the corresponding three DSCs. Also, VI was found to be unaffected by the number of experts used for the validation. VI may be used for a range of validation studies such as an evaluation of algorithmic segmentation methods.