Skip to main content
. Author manuscript; available in PMC: 2022 Jun 1.
Published in final edited form as: Artif Intell Med. 2021 Apr 23;116:102078. doi: 10.1016/j.artmed.2021.102078

Table 7.

Results of the experiment to investigate the impact of dataset size in the source and target domains on the effect of transfer learning. In this experiment, the source task is whole brain segmentation in the Whole-Brain-MRI dataset, and the target task is brain cortical plate segmentation in the CP-newborn dataset. For each setting of the number of images in the target domain, there were no statistically significant differences in the results obtained with different numbers of images in the source domain (p = 0.05).

number of images in the target domain number of images in the source domain DSC HD95 (mm)

50 0 (R.I.) 0.888 ± 0.036 0.841 ± 0.092
50 0.906 ± 0.039 0.837 ± 0.087
250 0.906 ± 0.035 0.837 ± 0.082
2500 0.908 ± 0.034 0.833 ± 0.084

150 0 (R.I.) 0.901 ± 0.035 0.837 ± 0.080
50 0.904 ± 0.036 0.840 ± 0.081
250 0.908 ± 0.038 0.836 ± 0.077
2500 0.910 ± 0.034 0.830 ± 0.066

450 0 (R.I.) 0.926 ± 0.029 0.810 ± 0.068
50 0.926 ± 0.033 0.820 ± 0.071
250 0.926 ± 0.030 0.813 ± 0.072
2500 0.930 ± 0.031 0.805 ± 0.061