Table 2.
WMH lesion segmentation results for the pre-trained baseline model without domain adaptation, transductive domain adaptation (TDA), and unsupervised LST method for four different sites: (1) source WMH 2017 and VH target; (2) source VH to Singapore; (3) source VH to UMC Utrecht; and (4) source VH to VU Amsterdam.
| WMH 2017 to VH (3T Siemens TrioTim) | VH to Singapore (3T Siemens TrioTim) | ||||||
|---|---|---|---|---|---|---|---|
| Baseline | TDA | LST | Baseline | TDA | LST | ||
| DSC | 0.478 ± 0.229 | 0.536 ± 0.232* | 0.410 ± 0.232 | DSC | 0.636 ± 0.176 | 0.703 ± 0.198* | 0.651 ± 0.176 |
| TPR | 0.735 ± 0.208 | 0.544 ± 0.231 | 0.319 ± 0.210 | TPR | 0.314 ± 0.089 | 0.451 ± 0.106 | 0.148 ± 0.092 |
| FPR | 0.611 ± 0.226 | 0.480 ± 0.256 | 0.477 ± 0.273 | FPR | 0.211 ± 0.186 | 0.469 ± 0.197 | 0.510 ± 0.153 |
| F-score | 0.270 ± 0.186 | 0.308 ± 0.187* | 0.160 ± 0.140 | F-score | 0.265 ± 0.102 | 0.289 ± 0.118 | 0.106 ± 0.067 |
| VH to UMC Utrecht (3T Philips Achieva) | VH to VU Amsterdam (3T GE Signa) | ||||||
| Baseline | TDA | LST | Baseline | TDA | LST | ||
| DSC | 0.587 ± 0.203 | 0.624 ± 0.210* | 0.620 ± 0.201 | DSC | 0.504 ± 0.148 | 0.602 ± 0.135* | 0.581 ± 0.155 |
| TPR | 0.464 ± 0.107 | 0.464 ± 0.148 | 0.250 ± 0.130 | TPR | 0.478 ± 0.114 | 0.483 ± 0.106 | 0.290 ± 0.105 |
| FPR | 0.279 ± 0.151 | 0.319 ± 0.175 | 0.352 ± 0.221 | FPR | 0.284 ± 0.155 | 0.298 ± 0.184 | 0.358 ± 0.161 |
| F-score | 0.316 ± 0.103 | 0.318 ± 0.111 | 0.181 ± 0.091 | F-score | 0.300 ± 0.108 | 0.341 ± 0.126* | 0.213 ± 0.095 |
DSC, dice similarity coefficient; TPR, true positive rate; FPR, false positive rate. Highest DSC and F-scores are shown in bold. Statistically significant improvements from baseline are indicated with “*”.