Table 2.
Mean absolute differences and standard deviation in pixel over all A-Scans in the test dataset.
| Control | AMD | |||||
|---|---|---|---|---|---|---|
| ILM | IBRPE | BM | ILM | IBRPE | BM | |
| Shah et al.12 | 1.04 ± 0.07 | 1.19 ± 0.18 | 1.54 ± 0.31 | 1.15 ± 0.25 | 1.88 ± 0.57 | 1.81 ± 0.56 |
| Liefers et al.13 | 0.840 | 1.280 | 1.227 | 1.055 | 1.568 | 1.858 |
| Asgari et al.5 | 0.65 ± 0.06 | 1.06 ± 0.12 | 0.9 ± 0.08 | 0.88 ± 0.09 | 1.23 ± 0.11 | 1.15 ± 0.1 |
| Model A (no layer order) | 0.39 ± 0.42 | 0.72 ± 0.39 | 0.85 ± 0.69 | 0.48 ± 0.75 | 0.77 ± 0.57 | 1.38 ± 3.02 |
| Model B (no attention) | 0.77 ± 1.45 | 0.76 ± 0.39 | 0.92 ± 1.24 | 0.91 ± 2.50 | 0.78 ± 0.59 | 1.22 ± 0.97 |
| Model C (no shortcus) | 0.33 ± 0.40 | 0.76 ± 0.48 | 0.86 ± 1.16 | 0.37 ± 0.61 | 0.79 ± 0.72 | 1.15 ± 0.89 |
| Final | 0.31 ± 0.31 | 0.71 ± 0.41 | 0.74 ± 0.47 | 0.37 ± 0.67 | 0.72 ± 0.66 | 1.08 ± 0.70 |
| Final BCI 95% | [0.26, 0.35] | [0.64, 0.79] | [0.67, 0.81] | [0.27, 0.47] | [0.61, 0.83] | [0.95, 1.20] |
Results are shown for three previous works, our final model, and 4 ablation studies where our model was trained without guaranteed order (A), without the self-attention block (B), and without shortcut connections (C). The 95% BCI was calculated from 10,000 resamples as above.
Best values are in [bold].