Table 1.
Mean performance (in %) for the different models on all holdout data sets of cross validation.
| Classifier | Skull stripping | Registration | Balanced accuracy | Sensitivity | Specificity | AUC |
|---|---|---|---|---|---|---|
| CNN | No | – | 71.26 ± 2.86% | 55.55 ± 7.51% | 86.96 ± 3.95% | 0.75 ± 0.02 |
| Lin. | 74.27 ± 3.83% | 63.13 ± 9.05% | 85.40 ± 6.45% | 0.80 ± 0.05 | ||
| Nonlin. | 77.61 ± 4.44% | 64.79 ± 5.02% | 90.43 ± 5.19% | 0.85 ± 0.06 | ||
| CNN | Yes | – | 77.66 ± 4.39% | 69.70 ± 7.65% | 85.63 ± 4.06% | 0.83 ± 0.05 |
| Lin. | 79.45 ± 3.34% | 76.87 ± 4.81% | 82.03 ± 6.23% | 0.86 ± 0.05 | ||
| Nonlin. | 82.13 ± 5.08% | 73.47 ± 7.89% | 90.78 ± 4.92% | 0.88 ± 0.05 | ||
| CNN+ | No | – | 80.66 ± 4.80% | 74.95 ± 7.85% | 86.36 ± 2.85% | 0.88 ± 0.04 |
| Lin. | 86.19 ± 6.01% | 79.73 ± 10.72% | 92.66 ± 3.73% | 0.92 ± 0.04 | ||
| Nonlin. | 83.50 ± 5.90% | 77.16 ± 8.95% | 89.83 ± 4.49% | 0.90 ± 0.04 | ||
| Logistic regression* | Yes | Lin.** | 82.00 ± 4.25% | 80.57 ± 7.16% | 83.43 ± 2.45% | 0.90 ± 0.04 |
Highest values per column are highlighted in bold.
AUC area under the curve of the receiver operating characteristics.
*Logistic regression by FSL-SIENAX.
**Linear registration is applied during FSL-SIENAX processing to obtain scaling factor.