Using u-net, all predicted values ​are learned as zero

Question

Basic u-net structure is used. shape is (112,112,1).

If it is a problem with the data set, many of the pixes are zero.

Therefore, it seems that the model comes out