Figure 9.

The results of brain extraction using U-net. For training, (a) 48 dHCP dataset, (b) 44 dHCP + 4 from our dataset, and (c) 25 sets from our dataset was used. The brain extraction network trained with each configuration was used to test dHCP and our infant data, as presented in the upper and lower rows, respectively. As demonstrated, the network performs well if the training and test datasets exhibit similar characteristics. When heterogeneous datasets were used for training and testing, the resulting masks tend to be less accurate.