Table 2.
Comparison of results for the healthy dataset.
| Network architecture | Conv/layer | Accuracy (SD) [%] |
Mean epoch train time (SD) [s] |
Median image evaluation time [msec] | Parameters [] |
|---|---|---|---|---|---|
| Vanilla U-Net | 2 | 99.10 (0.01) | 41.46 (0.60) | 128 | 1.95 |
| 3 | 99.20 (0.05) | 54.33 (0.34) | 138 | 2.93 | |
| Dense U-Net | 2 | 99.08 (0.02) | 46.36 (0.14) | 133 | 2.71 |
| 3 | 99.20 (0.05) | 72.88 (0.02) | 155 | 5.44 | |
| Attention U-Net | 2 | 99.12 (0.05) | 53.00 (0.85) | 137 | 1.99 |
| 3 | 99.23 (0.01) | 66.65 (0.52) | 148 | 2.98 | |
| SE U-Net | 2 | 99.08 (0.02) | 57.92 (0.95) | 143 | 2.06 |
| 3 | 99.20 (0.06) | 70.88 (0.27) | 155 | 3.04 | |
| Residual U-Net | 2 | 99.10 (0.04) | 42.61 (0.07) | 127 | 1.95 |
| 3 | 99.23 (0.02) | 55.50 (0.12) | 139 | 2.93 | |
| R2 U-Net | 2 | 99.24 (0.02) | 65.01 (1.04) | 156 | 2.00 |
| 3 | 99.28 (0.01) | 103.59 (0.23) | 186 | 3.00 | |
| U-Net + + | 99.12 (0.02) | 93.23 (1.38) | 165 | 2.05 | |
| Inception U-Net | 99.10 (0.04) | 108.31 (0.04) | 182 | 4.62 | |
(SD: standard deviation, s: seconds, SE: squeeze + excite, R2: recurrent-residual). The best and poorest results for each are annotated in bold and underline text respectively.