TABLE I.
Mean and standard deviation of the scores for different algorithms on LPBA40 and OASIS datasets. The results show that our algorithm increased both sensitivity and specificity and resulted in highest Dice scores among all widely-used tools and the recent PCNN method [12].
| LPBA40 | OASIS | |||||
|---|---|---|---|---|---|---|
| Method | Dice | Sensitivity | Specificity | Dice | Sensitivity | Specificity |
| Auto-U-net | 97.73 (±0.003) | 98.31 (±0.006) | 99.48 (±0.001) | 97.62 (±0.01) | 98.66 (±0.01) | 98.77 (±0.01) |
| U-net | 96.79 (±0.004) | 97.22 (±0.01) | 99.34 (±0.002) | 96.22 (±0.006) | 97.29 (±0.01) | 98.27 (±0.007) |
| Auto-2.5D-CNN | 97.66 (±0.01) | 98.25 (±0.01) | 99.47 (±0.002) | 96.06 (±0.007) | 96.21 (±0.01) | 98.56 (±0.006) |
| 2.5D-CNN | 97.17 (±0.005) | 98.52 (±0.01) | 99.24 (±0.002) | 95.61 (±0.007) | 96.3 (±0.01) | 98.20 (±0.01) |
| PCNN | 96.96 (±0.01) | 97.46 (±0.01) | 99.41 (±0.003) | 95.02 (±0.01) | 92.40 (±0.03) | 99.28 (±0.004) |
| BET | 94.57 (±0.02) | 98.52 (±0.005) | 98.22 (±0.01) | 93.44 (±0.03) | 93.41 (±0.04) | 97.70 (±0.02) |
| Robex | 95.40 (±0.04) | 94.25 (±0.05) | 99.43 (±0.004) | 95.33 (±0.01) | 92.97 (±0.02) | 99.21 (±0.004) |
| 3dSkullStrip | 92.99 (±0.03) | 96.95 (±0.01) | 97.87 (±0.01) | 92.77 (±0.01) | 94.44 (±0.04) | 96.82 (±0.01) |
| HWA | 92.41 (±0.007) | 99.99 (±0.0001) | 97.07 (±0.004) | 94.06 (±0.01) | 98.06 (±0.01) | 96.34 (±0.01) |