TABLE III:
Comparison of different semi-supervised learning frameworks trained with/without our SCO-SSL method for prostate ultrasound segmentation. Bold values indicate the best performance. Our SCO-SSL method can significantly improve the performance of most of the semi-supervised learning frameworks, except those underlined entries (p>0.05).
| Methods | UCLA dataset labeled/unlabeled = 194/701 |
NIH dataset labeled/unlabeled = 180/347 |
||||||
|---|---|---|---|---|---|---|---|---|
| DSC [%] |
ASD [mm] |
ASD-shadow [mm] |
HD [mm] |
DSC [%] |
ASD [mm] |
ASD-shadow [mm] |
HD [mm] |
|
| Π-model (2017) [36] | 91.39 (2.54) | 1.04 (0.36) | 1.21 (0.59) | 6.28 (2.34) | 89.27 (5.18) | 1.35 (0.79) | 1.96 (1.44) | 7.33 (3.86) |
| + SCO-SSL | 91.47 (2.54) | 1.04 (0.34) | 1.21 (0.52) | 6.32(2.11) | 89.53 (5.94) | 1.35(1.39) | 1.96(1.96) | 7.40(3.98) |
| Temporal ensembling (2017) [36] | 91.11(2.70) | 1.09(0.38) | 1.29(0.64) | 6.95(2.77) | 88.72(4.28) | 1.44(0.78) | 2.15 (1.70) | 8.14(4.03) |
| + SCO-SSL | 91.61 (2.37) | 1.02 (0.34) | 1.23 (0.55) | 6.30 (2.28) | 89.63 (5.49) | 1.32 (1.03) | 1.96 (2.01) | 6.94 (3.58) |
| Mean-teacher (2017) [37] | 91.36(2.56) | 1.05(0.36) | 1.26(0.56) | 6.48(2.34) | 89.55(4.19) | 1.30(0.66) | 1.87(1.23) | 7.12(3.32) |
| + SCO-SSL | 91.60 (2.37) | 1.02 (0.34) | 1.22 (0.52) | 6.37 (2.36) | 90.12 (3.61) | 1.23 (0.63) | 1.80 (1.18) | 6.65 (2.89) |
| UA-MT (2019) [38] | 91.29(2.68) | 1.06(0.37) | 1.27(0.57) | 6.64(2.51) | 89.54(3.75) | 1.31(0.65) | 1.88(1.23) | 7.04(3.20) |
| + SCO-SSL | 91.80 (2.42) | 0.99 (0.32) | 1.17 (0.51) | 6.12 (2.08) | 90.07 (3.74) | 1.23 (0.61) | 1.78 (1.14) | 6.61 (2.87) |
| TC-MT (2020) [41] | 91.13(2.74) | 1.08(0.39) | 1.31(0.58) | 6.91(2.75) | 89.21(4.88) | 1.36(0.74) | 1.96(1.40) | 7.69(3.75) |
| + SCO-SSL | 91.48 (2.41) | 1.04 (0.35) | 1.25 (0.56) | 6.44 (2.45) | 89.94 (3.53) | 1.25 (0.62) | 1.85 (1.25) | 6.92 (2.96) |