Table 4 Ablation study of label correction task.(%)
From: Self-relabeling for noise-tolerant retina vessel segmentation through label reliability estimation
Dataset | DRIVE(R) | STARE(VK) | CHASE | ||||
|---|---|---|---|---|---|---|---|
Group | Method | F1 | PR | F1 | PR | F1 | PR |
\(\mathrm {LV}\)-1 | w/o TML | 78.5 | 87.5 | 79.0 | 87.4 | 85.5 | 93.9 |
w \(\mathbf{S} ^j\) | 79.3 | 87.7 | 78.9 | 87.2 | 85.5 | 93.8 | |
w \(\widetilde{\mathbf{S }}^j\) | 79.8 | 88.1 | 79.5 | 87.7 | 86.2 | 94.2 | |
Ours | 79.6 | 88.6 | 79.9 | 88.4 | 85.9 | 94.1 | |
\(\mathrm {LV}\)-2 | w/o TML | 76.4 | 85.9 | 78.0 | 86.5 | 83.6 | 92.2 |
w \(\mathbf{S} ^j\) | 76.7 | 85.7 | 78.3 | 86.7 | 83.8 | 92.4 | |
w \(\widetilde{\mathbf{S }}^j\) | 77.1 | 84.7 | 77.9 | 87.0 | 84.5 | 92.6 | |
Ours | 77.6 | 87.1 | 78.9 | 87.3 | 84.3 | 92.6 | |
\(\mathrm {LV}\)-3 | w/o TML | 75.6 | 84.0 | 76.9 | 85.0 | 82.4 | 91.0 |
w \(\mathbf{S} ^j\) | 74.4 | 82.9 | 76.0 | 84.4 | 81.3 | 89.0 | |
w \(\widetilde{\mathbf{S }}^j\) | 77.1 | 84.7 | 77.9 | 87.0 | 84.5 | 92.6 | |
Ours | 77.0 | 86.4 | 77.6 | 86.4 | 83.1 | 91.4 | |
Pseudo | w/o TML | 79.9 | 88.0 | 76.2 | 83.9 | / | / |
w \(\mathbf{S} ^j\) | 80.0 | 88.0 | 76.3 | 84.0 | / | / | |
w \(\widetilde{\mathbf{S }}^j\) | 80.1 | 87.9 | 76.7 | 84.8 | / | / | |
Ours | 80.3 | 88.5 | 76.5 | 84.2 | / | / | |
Manual | w/o TML | 82.6 | 90.5 | 76.2 | 83.4 | 82.7 | 90.6 |
w \(\mathbf{S} ^j\) | 82.5 | 90.9 | 75.8 | 83.1 | 82.5 | 90.4 | |
w \(\widetilde{\mathbf{S }}^j\) | 82.6 | 90.8 | 77.6 | 84.7 | 82.8 | 90.1 | |
Ours | 83.0 | 91.3 | 77.8 | 84.3 | 83.3 | 91.1 | |