Table 7.
Mean (standard deviation) pixel accuracy (Acc.), precision (Pr.), recall (Rc.), and specificity (Sp.) overall performance for the Mask R-CNN, U-Net, FCN and DeeplabV3 methods for the six retinal regions (ILM-RPE, NFL-IPL, IPL-OPL, OPL-ELM, ELM-ISOS, ISOS-RPE) and retina/choroid segmentation. Each value represents the mean of four independent runs.
| Metric | Mask R-CNN | U-Net | FCN | DeeplabV3 |
|---|---|---|---|---|
| Retinal regions | ||||
| Acc. | 99.64 (0.01) | 99.74 (0.01) | 99.67 (0.01) | 99.43 (0.01) |
| Pr. | 94.22 (0.14) | 95.60 (0.03) | 93.14 (0.02) | 98.23 (0.02) |
| Rc. | 90.76 (0.13) | 94.78 (0.05) | 93.37 (0.04) | 98.28 (0.01) |
| Sp. | 99.87 (0.01) | 99.88 (0.01) | 99.83 (0.01) | 99.66 (0.04) |
| Retina/Choroid | ||||
| Acc. | 99.25 (0.04) | 99.55 (0.01) | 99.51 (0.01) | 99.43 (0.01) |
| Pr. | 99.08 (0.08) | 98.51 (0.24) | 98.27 (0.07) | 99.34 (0.02) |
| Rc. | 95.84 (0.29) | 98.38 (0.20) | 98.24 (0.03) | 99.33 (0.01) |
| Sp. | 99.85 (0.01) | 99.75 (0.04) | 99.70 (0.01) | 99.51 (0.01) |