Validation of Segmentation Methods

New graph cut segmentation and manual segmentation were analyzed in two sub-studies. In a full analysis study, all 48 bones were segmented with both approaches by a single operator. In the inter-operator reproducibility study, a subset of 12 left femur bones were segmented with both approaches by two additional operators (total of three operators).

In order to compare interactive graph cut segmentation to the manual gold standard, voxel masks were compared, and the absolute maximum difference (mm) (Hausdorff distance, Insight Toolkit, www.itk.org) and mean overlap (Dice coefficient, Insight Toolkit, www.itk.org) were computed for each sample. Additionally, differences in segmentations were assessed by computing the average, minimum and maximum signed surface-to-surface differences [mm] (Visualization Toolkit, www.vtk.org) of smoothed segmentation surfaces. These surface-to-surface differences from the full analysis study were subsequently compared to differences in FE results. In addition to comparing graph cut to manual segmentations, in the inter-operator reproducibility study, these metrics were calculated between graph cut segmentations and manual segmentations from the three operators in pair-wise manner.