On the usage of average Hausdorff distance for segmentation performance assessment: hidden error when used for ranking

摘要

Flow chart of the error simulation framework and correlation analysis. a The ground truth was created using a U-Net deep learning architecture and subsequently manually corrected. b The voxels in the errors are added or subtracted from the ground truth depending on whether the error is a false positive or false negative error. b1 Error introducing false positive voxels (green) in the skull area. b2 Error in which false-negative voxels (white) in the M3 segment of the middle cerebral artery are missing. c1 False-positive voxels in the skull area are added to the ground truth to create a simulated segmentation. c2 The error simulation framework allows the random combination of manually created errors to create simulated segmentations containing multiple errors. This simulated segmentation was created by combining seven errors. d The ten simulated segmentations in the set have an increasing number of errors. e The simulated segmentations are ranked from best to worst using the average Hausdorff distance and balanced average Hausdorff distance values, respectively. f Lastly, the correlation between the rankings are measured by the Kendall rank correlation coefficient. The process is repeated using 20 sets of simulations for each patient