Fig. 2.

Multiple network structures for comparison. The plain U-Net structure (left) is used in stage one and two as the baseline. Top right shows an example of residual block, which adds the previous filter to the latter by elementwise addition. Bottom right shows an example of a four-layer dense block, which concatenates each layer to every other layer. Either the residual or the dense block replaces the conventional convolutional block in each resolution level (as shown in red box) to build the corresponding networks.