Table 2:
Comparison of StackGen-Net with other WMH detection techniques
| Test Cohort 1 |
Test Cohort 2 |
|||||||
|---|---|---|---|---|---|---|---|---|
| UNET2D | DeepMedic | UNET2D WS-E | StackGen-Net | UNET2D | DeepMedic | UNET2D WS-E | StackGen-Net | |
| Dice (F1-P) | 0.43 [SD, 0.17] |
0.62 [SD, 0.09] |
0.67 [SD, 0.09] |
0.76 [SD, 0.07] |
0.27 [SD, 0.20] |
0.58 [SD, 0.15] |
0.66 [SD, 0.17] |
0.76 [SD, 0.09] |
| Precision-P | 0.72 [SD, 0.19] |
0.63 [SD, 0.13] |
0.72 [SD, 0.15] |
0.73 [SD, 0.11] |
0.73 [SD, 0.32] |
0.66 [SD, 0.22] |
0.69 [SD, 0.23] |
0.77 [SD, 0.11] |
| Recall-P | 0.32 [SD, 0.19] |
0.63 [SD, 0.18] |
0.64 [SD, 0.07] |
0.79 [SD, 0.1] |
0.18 [SD, 0.16] |
0.53 [SD, 0.13] |
0.67 [SD, 0.12] |
0.75 [SD, 0.09] |
| Precision-L | 0.60 [SD, 0.20] |
0.47 [SD, 0.23] |
0.69 [SD, 0.18] |
0.75 [SD, 0.11] |
0.72 [SD, 0.24] |
0.43 [SD, 0.23] |
0.60 [SD, 0.23] |
0.84 [SD, 0.14] |
| Recall-L | 0.37 [SD, 0.09] |
0.86 [SD, 0.09] |
0.79 [SD, 0.15] |
0.87 [SD, 0.08] |
0.26 [SD, 0.14] |
0.71 [SD, 0.10] |
0.74 [SD, 0.13] |
0.67 [SD, 0.13] |
| F1-L | 0.44 [SD, 0.10] |
0.54 [SD, 0.11] |
0.71 [SD, 0.09] |
0.80 [SD, 0.09] |
0.37 [SD, 0.16] |
0.50 [SD, 0.20] |
0.63 [SD, 0.15] |
0.73 [SD, 0.11] |
| |VD|(%) | 54.4 [SD, 22.1] |
26.9 [SD, 20.0] |
17.6 [SD, 11.2] |
12.3 [SD, 12.7] |
77.4 [SD, 16.5] |
30.6 [SD, 18.6] |
37.6 [SD, 51.5] |
13.7 [SD, 9.7] |
| HD95 | 19.5 [SD, 8.6] |
15.9 [SD, 16.1] |
10.8 [SD, 6.7] |
5.27 [SD, 3.15] |
30.6 [SD, 20.9] |
21.8 [SD, 22.9] |
19.5 [SD, 18.8] |
17.1 [SD, 21.0] |
| AUC | 0.53 [SD, 0.21] |
0.66 [SD, 0.12] |
0.61 [SD, 0.11] |
0.84 [SD, 0.07] |
0.54 [SD, 0.28] |
0.60 [SD, 0.20] |
0.60 [SD, 0.20] |
0.84 [SD, 0.10] |
Note:—HD95 indicates modified Hausdorff distance (mm); P, pixel; L, lesion; |VD| = absolute volume difference.