Fig. 2.

Uncertainty of semantic segmentation, pixel-wise regression and classification can be estimated with InstantDL. a The distribution of average pixel uncertainty u for 162 images of a semantic segmentation task [13]. The distribution is approximately bi-modal. b Three exemplary semantic segmentations from the data visualized in a with InstantDLs pixel uncertainty estimation. Correct predictions correspond to a low average pixel uncertainty u (top and middle row), while a high average pixel uncertainty indicates failed segmentations (bottom row). Regions with ambiguous predictions are indicated by high pixel uncertainty (right column). c For the prediction of white blood cells classes [6], classification uncertainty indicates incorrect predictions. d The distributions of classification uncertainty for correct and false predictions differ significantly (p value < 0.001, Mann–Whitney rank test)