Fig. 6.

Interactions by real human users with between 1 and 4 years experience working with glioblastomas for whole tumor region. Users were first asked to annotate without instructions, then used the UNCERTAIN method in separate runs. For comparison, the simulated UNCERTAIN (along with $1\sigma$ standard deviation) as well as MISCLASS annotations are given. The latter is methodically most similar to the users’ intuitive approach. Intuitive annotations perform better than annotations in uncertain regions, likely because the majority of annotations users provide are corrective. Overall scores are lower than what was achieved in the simulations, likely because users do not fully satisfy our assumption that they possess knowledge of the correct segmentation. Because the number of data collected from real users was small in number, the comparison to simulated interactions should be understood qualitatively. Inlay: the distribution of scribble lengths.