Table 2.
Comparing the properties and performance of different segmentation networks. Numbers in the parenthesis show the 95% confidence interval. Best performance in each metric is emphasized by bold text. Missed cases shows the percentage of cases with a Dice = 0 in the test set.
| Model | Number of Trainable Parameters | Dice | Surface Dice | Number of Non-Overlapping Connected Components | Volume ICC | 95th Percentile HD | Missed Cases |
|---|---|---|---|---|---|---|---|
| U-Net | 82M | 0.59 (0.55, 0.64) |
0.75 (0.69, 0.80) |
2.5 (1.5, 3.5) |
0.70 | 10.46 (7.33, 13.58) |
6% |
| Concatenate | 115M | 0.60 (0.55, 0.65) |
0.76 (0.71, 0.82) |
0.6 (0.3, 0.9) |
0.68 | 5.89 (4.36, 7.42) |
8% |
| Weighted-sum | 97M |
0.62 (0.58, 0.66) |
0.78 (0.73, 0.83) |
0.7 (0.4, 0.9) |
0.60 | 5.88 (4.54, 7.22) |
4% |
| Add | 96M | 0.53 (0.48, 0.58) |
0.69 (0.63, 0.76) |
0.3 (0.1, 0.4) |
0.49 |
5.67 (4.30, 7.04) |
11% |
| U-Net with no CTA input | 82M | 0.38 (0.32, 0.45) |
0.51 (0.43, 0.58) |
0.2 (0.1, 0.3) |
0.41 | 7.75 (5.25, 10.26) |
26% |