Label-Set Loss Functions for Partial Supervision: Application to Fetal Brain 3D MRI Parcellation

Deep neural networks have increased the accuracy of automatic segmentation, however their depends on availability a large number fully segmented images. Methods to train deep using images for which some, but not all, regions interest are necessary make better use partially annotated datasets. In this paper, we propose first axiomatic definition label-set loss functions that can handle We prove there is one and only method convert classical function into proper function. Our theory also allows us define leaf-Dice loss, generalisation Dice particularly suited partial supervision with missing labels. Using set new state art in supervised learning fetal brain 3D MRI segmentation. achieve network able segment white matter, ventricles, cerebellum, extra-ventricular CSF, cortical gray brainstem, corpus callosum based anatomically normal fetuses or open spina bifida. implementation proposed available at https://github.com/LucasFidon/label-set-loss-functions.