Skip to main content
. 2022 Mar 15;41(4):706–717. doi: 10.14366/usg.21214

Table 3.

Comparison of segmentation performance with other U-Net variants

Architecture Precision Recall DSC HD
HM70A wrist dataset
 U-Net 0.916 0.895 0.897 5.476
 U-Net++ 0.914 0.884 0.889 5.577
 Attention U-Net 0.918 0.890 0.896 5.442
 MultiRes U-Net 0.934a) 0.869 0.893 5.475
 Proposed 0.927 0.924a) 0.923a) 5.158a)
HM70A forearm dataset
 U-Net 0.763 0.738 0.719 5.441
 U-Net++ 0.790 0.712 0.713 5.487
 Attention U-Net 0.793 0.742 0.744 5.283
 MultiRes U-Net 0.727 0.666 0.653 6.050
 Proposed 0.793a) 0.760a) 0.761a) 5.206a)
miniSONO wrist dataset
 U-Net 0.859 0.829 0.830 5.430
 U-Net++ 0.882 0.827 0.839 5.243
 Attention U-Net 0.857 0.816 0.824 5.449
 MultiRes U-Net 0.897 0.823 0.848 5.141
 Proposed 0.903a) 0.898a) 0.897a) 4.966a)
miniSONO forearm dataset
 U-Net 0.873 0.828 0.838 4.661
 U-Net++ 0.857 0.805 0.812 4.844
 Attention U-Net 0.864 0.820 0.828 4.738
 MultiRes U-Net 0.868 0.793 0.813 4.774
 Proposed 0.878a) 0.850a) 0.858a) 4.527a)

DSC, dice similarity coefficient; HD, Hausdorff distance.

a)

The best performance of each metric.