Figure 1
Diagram of a 2D four-layer tunable U-Net congruent with input data of cin channels and spatial dimensions m and n. Among the user-defined hyperparameters on display are the initial base channels cb and the channel growth factor r, both of which control the size of the network and thus its potential expressive power. The hidden growth rate rh is set to 1 for simplicity. We note that DLSIA easily accommodates volumetric data by simply replacing all convolutions (and associated layer normalization) with their 3D counterparts. |