Fig. 21.

Comparing segmentation-performance prediction with an earlier model [3]: predicting dice similarity for large database sizes using small databases for (a) subcortical structures in brain MR images. Box-whisker plots show the empirically computed DSC values DSC(M_i) for M_i ∈ {10, 15, 20, …} for the number of templates fixed to 8, independent of database size M_i. Variance depicted by the whiskers comes from the bootstrap samples of small databases ${\left\{b_l^{\tilde{N}}\right\}}_{l=1}^L$ available for analysis (we use L = 20). Solid lines and the associated error bars show the mean and standard deviation, respectively, of the fitted DSC values, using the larger databases $b_l^N$; the error bars come from the bootstrap samples of databases. Dashed lines and associated error bars show the mean and standard deviation, respectively, of the fitted DSC values using small databases $b_l^{\tilde{N}}$; the error bars come from the bootstrap samples of databases.