Table 3.
Dice and Hausdorff distance of different methods on validation set of BraTS 2018 (mean ± std).
| Dice | Hausdorff (mm) | |||||
|---|---|---|---|---|---|---|
| ET | WT | TC | ET | WT | TC | |
| 3D UNet | 0.734 ± 0.284 | 0.864 ± 0.146 | 0.766 ± 0.230 | 9.37 ± 22.95 | 12.00 ± 21.22 | 10.37 ± 13.47 |
| 3D UNet + TTA | 0.754 ± 0.263 | 0.873 ± 0.125 | 0.783 ± 0.168 | 4.53 ± 9.60 | 5.90 ± 6.80 | 8.03 ± 10.31 |
| Multi-class WNet | 0.757 ± 0.257 | 0.890 ± 0.089 | 0.725 ± 0.245 | 4.24 ± 7.97 | 4.99 ± 6.53 | 12.13 ± 13.41 |
| Multi-class WNet + TTA | 0.771 ± 0.242 | 0.896 ± 0.071 | 0.730 ± 0.255 | 4.44 ± 8.20 | 4.92 ± 6.42 | 11.13 ± 13.46 |
| Cascaded networks | 0.792 ± 0.233 | 0.903 ± 0.057 | 0.854 ± 0.142 | 3.34 ± 4.15 | 5.38 ± 9.31 | 6.61 ± 8.55 |
| Cascaded networks + TTA | 0.797 ± 0.229 | 0.902 ± 0.056 | 0.858 ± 0.139 | 3.13 ± 3.78 | 6.18 ± 9.53 | 6.37 ± 8.19 |
| Cascaded networks + TTA + CRF0 | 0.803 ± 0.228 | 0.905 ± 0.056 | 0.862 ± 0.136 | 3.09 ± 3.75 | 5.97 ± 8.22 | 6.25 ± 7.87 |
| Cascaded networks + TTA + CRF1 | 0.807 ± 0.225 | 0.908 ± 0.054 | 0.869 ± 0.126 | 3.01 ± 3.69 | 5.86 ± 8.16 | 6.09 ± 7.74 |
| Myronenko, 2018 (single model) | 0.815 | 0.904 | 0.860 | 3.80 | 4.48 | 8.28 |
| Myronenko, 2018 (ensemble) | 0.823 | 0.910 | 0.867 | 3.93 | 4.52 | 6.85 |
| Isensee et al., 2018 | 0.810 | 0.908 | 0.854 | 2.54 | 4.97 | 7.04 |
ET, enhancing tumor core; WT, whole tumor; TC, tumor core; TTA, test-time augmentation. CRF0: naive conditional random field for post-processing. CRF1: our uncertainty-aware conditional random field. Bold value shows the best performance.