Figure 2.

Model architectures. Overview of internal layers in 2D (A) and 3D U-Net (B) utilized to perform brain segmentation. A two or three-dimensional version of internal layers is used depending on slice-wise or patch-wise inputs. The output layer(s) depends on single or multi-task learning. For 2D U-Nets we used single-task learning, hence there was a single output layer. For 3D U-Nets, three tissue classes were trained at a time along with background hence there were four output layers.