Table 4.
Statistics of the compared methods on the test set (22 cases) based on the contours generated by Simultaneous Truth and Performance Level Estimation Algorithm. Those networks trained with data augmentation can achieve significantly better performance than those without data augmentation
| Methods | Modalities | With data‐augmentation | Without data‐augmentation | ||||
|---|---|---|---|---|---|---|---|
| Score | DSC | ASSD | Score | DSC | ASSD | ||
| 3D‐UNet (CELoss) | CT | 0.812 ± 0.115 | 0.780 ± 0.185 | 1.667 ± 1.863 | 0.752 ± 0.157 | 0.719 ± 0.207 | 3.154 ± 4.336 |
| PET | 0.846 ± 0.084 | 0.811 ± 0.133 | 1.127 ± 0.718 | 0.819 ± 0.078 | 0.779 ± 0.127 | 2.491 ± 2.809 | |
| DFCN‐CoSeg (CELoss) | CT | 0.850 ± 0.061 | 0.836 ± 0.095 | 0.895 ± 0.661 | 0.818 ± 0.084 | 0.806 ± 0.108 | 1.226 ± 0.987 |
| PET | 0.848 ± 0.064 | 0.823 ± 0.086 | 1.066 ± 0.660 | 0.793 ± 0.101 | 0.771 ± 0.116 | 1.993 ± 2.346 | |
| 3D‐UNet (DICELoss) | CT | 0.839 ± 0.085 | 0.811 ± 0.151 | 1.291 ± 1.313 | 0.759 ± 0.156 | 0.728 ± 0.204 | 3.431 ± 5.176 |
| PET | 0.832 ± 0.075 | 0.794 ± 0.111 | 1.229 ± 0.587 | 0.832 ± 0.076 | 0.808 ± 0.105 | 1.308 ± 0.750 | |
| DFCN‐CoSeg (DICELoss) | CT | 0.865 ± 0.034 | 0.861 ± 0.037 | 0.806 ± 0.605 | 0.811 ± 0.098 | 0.785 ± 0.152 | 2.632 ± 4.657 |
| PET | 0.853 ± 0.063 | 0.828 ± 0.087 | 1.079 ± 0.761 | 0.803 ± 0.087 | 0.778 ± 0.107 | 1.856 ± 1.976 | |