Table 7.
Results from ablation study of the DEP-GAN-2C’s regularisation terms tested using PM (see Eq. (4)). We calculated the prediction error of WMH change, volumetric agreement of WMH volume, and spatial agreement of WMH evolution, compared to the gold standard expert-delineated WMH masks (i.e., three-class DEM labels). “DSC” stands for similarity coefficient, “Int.” stands for intensity,“Vol.” stands for volumetric, “LoA” stands for limit of agreement, and “G” and “S” stand for percentage of subjects correctly predicted as having growing and shrinking WMH by DEP models. The best value for each learning approaches and evaluation measurements is written in bold.
| DEP-GAN-2C (PM) |
Grow | Shrink | Avg. [%] | Vol. Bias [ml] | Lower | Upper | Entire | Change | Stable | Shrink | Grow | Avg. ((St+ | ||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| λ1 (Int.) | λ2 (DSC) | λ3 (Vol.) | (G) [%] | (S) [%] | ((G+S)/2) | mean(std) | LoA [ml] | LoA [ml] | WMH | (C) | (St) | (Sr) | (Gr) | Sr+Gr)/3) |
| 0 | 0 | 0 | 64.29 | 85.19 | 74.74 | 3.03(7.65) | −11.9684 | 18.0372 | 0.6131 | 0.1667 | 0.6178 | 0.1045 | 0.0813 | 0.2679 |
| 0 | 0 | 100 | 65.31 | 79.63 | 72.47 | 2.28(8.16) | −13.7197 | 18.2747 | 0.6132 | 0.1749 | 0.6166 | 0.1009 | 0.0909 | 0.2695 |
| 0 | 1 | 0 | 50.00 | 83.33 | 66.67 | 4.32(8.18) | −11.7181 | 20.3473 | 0.6093 | 0.1919 | 0.6063 | 0.1366 | 0.0706 | 0.2712 |
| 100 | 0 | 0 | 57.14 | 83.33 | 70.24 | 3.79(7.83) | −11.5525 | 19.1234 | 0.6075 | 0.1827 | 0.6143 | 0.1312 | 0.0741 | 0.2732 |
| 0 | 1 | 100 | 67.35 | 75.93 | 71.64 | 2.37(8.50) | −14.2904 | 19.0237 | 0.6101 | 0.1889 | 0.6177 | 0.1203 | 0.0922 | 0.2767 |
| 100 | 1 | 0 | 58.16 | 77.78 | 67.97 | 2.23(8.85) | −15.1197 | 19.5748 | 0.6096 | 0.1912 | 0.6079 | 0.1209 | 0.0925 | 0.2738 |
| 100 | 0 | 100 | 57.14 | 88.89 | 73.02 | 4.51(8.15) | -11.4546 | 20.4778 | 0.6078 | 0.1993 | 0.5996 | 0.1446 | 0.0760 | 0.2734 |
| 100 | 1 | 100 | 56.12 | 81.48 | 68.80 | 3.46(8.26) | −12.7218 | 19.6500 | 0.6107 | 0.1801 | 0.6245 | 0.1216 | 0.0868 | 0.2776 |