Fig. 4.

The Pearson correlation coefficient is not necessarily an accurate indicator of quality prediction performance. Scatter plots of the predicted vs the observed DSCs are shown for: (A) all the candidate segmentations in the QCD framework, (B) U-net 7, and (C) the final segmentations selected by the QCD framework. The highest classification accuracy (ACC) of good (observed DSC $\ge 0.7$) and bad (observed DSC $<0.7$) segmentations is seen in (C) the final segmentations selected by the QCD framework (ACC=0.99), compared to (A) all the candidate segmentations (ACC=0.96) and (B) U-net 7 (ACC=0.94). Although high correlations were observed for (A) all the candidate segmentations ($r=0.92$) and (B) U-net 7 ($r=0.94$), a much weaker correlation was obtained for (C) the final QCD segmentations ($r=0.53$), which had a better segmentation performance (observed DSC between 0.59-0.95) and despite having the highest accuracy (ACC=0.99).