Table 6 MRGAM branch-weighting mechanisms on LUNA16 and Tianchi detection
From: GLANCE: continuous global-local exchange with consensus fusion for robust nodule segmentation
Weighting mechanism | LUNA16 detection | Tianchi detection | ||||
|---|---|---|---|---|---|---|
| Â | Pre. | Recall | F1 | Pre. | Recall | F1 |
Global+Local softmax (ours) | 92.3 | 94.1 | 93.2 | 95.0 | 93.6 | 94.3 |
Uniform (1/N) | 90.1 | 91.8 | 90.9 | 93.0 | 92.0 | 92.5 |
Local-SE only | 90.8 | 91.9 | 91.3 | 93.4 | 92.6 | 93.0 |
Static scalars (learned) | 90.5 | 91.7 | 91.1 | 93.2 | 92.4 | 92.8 |
1 × 1-Conv mixing (no gating) | 90.6 | 91.8 | 91.2 | 93.3 | 92.6 | 92.9 |
Scaled dot-product attention | 90.7 | 91.8 | 91.2 | 93.4 | 92.7 | 93.0 |
