Table 2.
IE-Vnet segmentation results on four test datasets D2-D5 (20 inner ears each) compared to manual ground-truth.
| Dataset | D2 | D3 | D4 | D5 | ||||
|---|---|---|---|---|---|---|---|---|
| Mean | SD | Mean | SD | Mean | SD | Mean | SD | |
| (A) Accuracy | ||||||||
| Dice | 0.906 | 0.012 | 0.895 | 0.021 | 0.904 | 0.014 | 0.894 | 0.026 |
| HDmax | 0.949 | 0.862 | 0.804 | 0.476 | 1.170 | 0.774 | 0.811 | 0.565 |
| SDmean | 0.020 | 0.003 | 0.023 | 0.005 | 0.021 | 0.003 | 0.023 | 0.006 |
| (B) Performance | ||||||||
| Cochlea | 4.950 | 0.224 | 4.900 | 0.308 | 4.900 | 0.308 | 0.950 | 0.224 |
| Sacculus | 5.000 | 0.000 | 5.000 | 0.000 | 4.950 | 0.224 | 5.000 | 0.000 |
| Utriculus | 5.000 | 0.000 | 5.000 | 0.000 | 5.000 | 0.000 | 5.000 | 0.000 |
| aSCC | 4.800 | 0.616 | 4.800 | 0.616 | 4.900 | 0.308 | 4.750 | 0.716 |
| pSCC | 4.900 | 0.447 | 4.900 | 0.447 | 4.850 | 0.671 | 4.900 | 0.308 |
| hSCC | 4.950 | 0.224 | 4.850 | 0.671 | 4.850 | 0.671 | 4.800 | 0.696 |
Segmentation accuracy (A) was measured by the quantitative metrics of Dice overlap coefficient (“Dice”), Hausdorff maximum surface distance (“HDmax,” in [mm]), and average surface distance (“SDmean,” in [mm]). Localized performance (B) issues within the inner ear were assessed using a semi-quantitative five-point Likert-type response scale for the cochlea, sacculus, utriculus, the anterior semi-circular canal (aSCC), posterior SCC (pSCC), and horizontal SCC (hSCC) respectively. A detailed description of the used categories can be found in Section 2.6.