TABLE II:
Inter-dataset evaluation of the fully-supervised methods. Bold values indicate the best result. Our method significantly outperformed most of the competing methods, except on those underlined entries (p>0.05).
| Methods | Train&val. on UCLA / test on NIH | Train&val. on NIH / test on UCLA | ||||||
|---|---|---|---|---|---|---|---|---|
| DSC [%] |
ASD [mm] |
ASD-shadow [mm] |
HD [mm] |
DSC [mm] |
ASD [mm] |
ASD-shadow [mm] |
HD [mm] |
|
| Radial-2.5D-UNet (2020) [20] | 63.10(16.41) | 4.14(1.68) | 5.21(2.13) | 14.59(4.67) | 75.41(8.04) | 3.56(1.10) | 4.16(1.89) | 17.67(5.18) |
| VNet (2016) [21] | 78.80 (10.55) | 3.09(1.75) | 4.99(3.49) | 16.88(7.47) | 54.86(13.00) | 9.21(3.45) | 12.01(5.03) | 35.01(9.25) |
| UNet (2015) [8] | 76.74(9.53) | 3.77(1.62) | 5.48(3.04) | 21.51(7.00) | 72.26(10.85) | 4.94(2.84) | 6.65(4.52) | 25.35(10.07) |
| nnUNet (2021) [18] | 74.42(13.27) | 4.42(2.72) | 7.80(5.70) | 22.27(11.09) | 74.47(11.74) | 4.52(2.72) | 6.60(4.70) | 23.47(9.67) |
| DAF-Net (2019) [11] | 79.97 (11.05) | 3.07(2.06) | 5.23(5.27) | 16.62(10.42) | 78.28 (10.57) | 3.56(2.73) | 4.76(4.65) | 19.37(10.65) |
| SCO-SSL (semi-supervised) | 75.58(11.14) | 4.15(2.17) | 5.26(3.65) | 23.03(8.96) | 80.93 (9.35) | 2.55 (2.06) | 3.23(2.53) | 13.30 (5.86) |
| SCO-SSL (full-supervised) | 79.51(10.06) | 2.71 (1.76) | 4.14 (2.48) | 12.97 (6.15) | 78.63(9.22) | 2.68(1.32) | 3.11 (1.89) | 13.89(4.44) |