Fig. 2.

Multi-task architecture for simultaneous segmentation and uncertainty prediction. All convolutions are 3 × 3 by the channel width, denoted in the diagram. The U-Net is complemented by an ‘arm’ which has residual blocks followed by max-pooling (maintaining 64 filters) until it reaches the output layer, where it returns an upper and lower bound. Dropout is enabled for methods that require MC-sampling where indicated in the diagram, with p = 0.5. The bounds are trained with the loss from Eq. 2.