Table 9.
Cross-dataset analysis on the full retinal tissue segmentation results, presenting the mean (standard deviation) mean signed and absolute boundary errors (in pixels) results. Three different training/testing combinations were included for each model, including: (i) Using only the original dataset (O.D./O.D.), (ii) training adding the AMD dataset images (O.D. + AMD), and testing only on O.D. and (iii) training adding the AMD dataset images (O.D. + AMD), and testing only on AMD. The O.D. dataset denotes original healthy dataset imaged with the Spectralis device and the AMD dataset denotes the pathology dataset imaged with the Bioptigen device.
| Data | Mask R-CNN | U-Net | FCN | DeeplabV3 | ||||
|---|---|---|---|---|---|---|---|---|
| Trained/ Tested | ILM | RPE | ILM | RPE | ILM | RPE | ILM | RPE |
| Mean signed error | ||||||||
| O.D./O.D. | −0.063 (0.133) | 0.125 (0.119) | −0.456 (0.040) | −0.447 (0.073) | 0.526 (0.036) | 0.461 (0.276) | 0.527 (0.034) | 0.376 (0.053) |
| O.D. + AMD/O.D. | 0.451 (0.270) | −0.169 (0.010) | 0.312 (0.007) | 0.121 (0.102) | 0.265 (0.148) | 0.199 (0.103) | 0.175 (0.084) | 0.151 (0.074) |
| O.D. + AMD/AMD | 1.299 (0.398) | −0.975 (0.810) | −0.034 (0.078) | 0.030 (0.134) | −0.093 (0.032) | −0.140 (0.044) | 0.035 (0.061) | −0.018 (0.140) |
| Mean absolute error | ||||||||
| O.D./O.D. | 0.960 (0.017) | 0.941 (0.017) | 0.601 (0.015) | 0.596 (0.041) | 0.744 (0.015) | 0.870 (0.050) | 0.736 (0.016) | 0.763 (0.020) |
| O.D. + AMD/O.D. | 1.250 (0.316) | 0.995 (0.187) | 0.629 (0.002) | 0.559 (0.052) | 0.507 (0.060) | 0.497 (0.037) | 0.482 (0.020) | 0.511 (0.026) |
| O.D. + AMD/AMD | 1.887 (0.308) | 2.434 (0.159) | 0.343 (0.019) | 0.408 (0.014) | 0.387 (0.011) | 0.523 (0.024) | 0.421 (0.004) | 0.557 (0.021) |