Vessel segmentation algorithm accuracy.

Comparison of the accuracy of unsupervised vessel segmentation algorithms as applied to the DRIVE image database. The hand segmented images from the first manual observer are used as the benchmark. True and false positive rates (TPR and FPR) are included where these were made available in the original papers. Note that assigning all pixels to the background – i.e. detecting no vessels at all – still achieves an accuracy score of 0.8727. On the other hand, a second manual observer achieved an accuracy of 0.9473. The accuracy of segmentation algorithms can therefore be expected to fall within this range; improving on the accuracy score of the second observer is not necessarily beneficial, since the choice of the first observer as the benchmark is arbitrary.