Table 3.
The five networks that achieved the best results were compared regarding parameters and computational time over ten-fold cross-validation (mean ± standard deviation)
| Model | Backbone | Optimizer | Number of parameters/million | Training time (s/epoch) | Segmentation time (ms/image) |
|---|---|---|---|---|---|
| CABM-HRNet | hrnetv2_w32 | Adam | 30.598 | 90.75 ± 0.825 | 12.75 ± 0.093 |
| HRNet | hrnetv2_w32 | Adam | 29.547 | 91.02 ± 0.559 | 12.65 ± 0.052 |
| PSPNet | ResNet50 | Adam | 2.377 | 108.56 ± 0.658 | 9.98 ± 0.076 |
| DeeplabV3 + | MobileNetv2 | Adam | 5.818 | 89.29 ± 0.612 | 8.62 ± 0.105 |
| U-Net | VGG | Adam | 24.892 | 138.11 ± 1.744 | 9.59 ± 0.070 |
CBAM has a lower overhead and computational load