From: Imaging segmentation mechanism for rectal tumors using improved U-Net
Method | Dice | MAP | MIoU | FWIoU |
---|---|---|---|---|
(a) Comparison results for each component | ||||
 Without ResNeSt | 0.923 | 0.825 | 0.776 | 0.781 |
 Without shape | 0.901 | 0.811 | 0.734 | 0.740 |
 Without PAM&CAM | 0.958 | 0.786 | 0.803 | 0.791 |
 Proposed U-Net | 0.987 | 0.946 | 0.897 | 0.899 |
(b) Comparison results for different attention mechanisms | ||||
 With SE | 0.935 | 0.755 | 0.645 | 0.611 |
 With GC | 0.949 | 0.809 | 0.774 | 0.740 |
 With CBAM | 0.961 | 0.902 | 0.812 | 0.827 |
 Proposed U-Net | 0.987 | 0.946 | 0.897 | 0.899 |
(c) Comparison of the different backbones used in the proposed U-Net network | ||||
 ResNet34 | 0.935 | 0.665 | 0.398 | 0.423 |
 SEResNeXt50 | 0.951 | 0.805 | 0.734 | 0.752 |
 SENet-154 | 0.958 | 0.911 | 0.860 | 0.854 |
 ResNeSt | 0.987 | 0.946 | 0.897 | 0.899 |
(d) Effect of the gate module | ||||
 Without a gate module | 0.973 | 0.922 | 0.855 | 0.861 |
 With the gate module | 0.987 | 0.946 | 0.897 | 0.899 |
(e) Results of comparison with existing advanced models | ||||
 DeepLabv3 [18] | 0.938 | 0.745 | 0.570 | 0.566 |
 U-Net++ [36] | 0.943 | 0.811 | 0.681 | 0.682 |
 U-Net+++ [37] | 0.925 | 0.707 | 0.550 | 0.554 |
 GSCNN [38] | 0.910 | 0.602 | 0.419 | 0.510 |
 ERFNet [39] | 0.946 | 0.843 | 0.473 | 0.473 |
 ET-Net [40] | 0.927 | 0.862 | 0.689 | 0.784 |
 Proposed U-Net | 0.987 | 0.946 | 0.897 | 0.899 |