Efforts are required to design Deep Learning models that are not only powerful, but also capable of expressing the certainty of their predictions.
We evaluate 3 state-of-the-art techniques for uncertainty quantification : Monte-Carlo Dropout, Deep Ensemble, and Heteroscedastic models. Evaluation is illustrated on a task of automatic segmentation of White-Matter Hyperintensities in T2-weighted FLAIR MRI sequences of Multiple-Sclerosis patients. Analysis is performed at 3 different scales : the voxel, the lesion, and the whole image. Results indicate the superiority of the Heteroscedastic approach, which ranked first in both the uncertainty and segmentation tasks.