Table 5 Ablation study results on LUNA16 and Tianchi detection
From: GLANCE: continuous global-local exchange with consensus fusion for robust nodule segmentation
Model variant | LUNA16 detection | Tianchi detection | ||||
|---|---|---|---|---|---|---|
Pre. | Recall | F1 | Pre. | Recall | F1 | |
Full GLANCE (GCT+MRGAM+CSCF) | 92.3 | 94.1 | 93.2 | 95.0 | 93.6 | 94.3 |
w/o CSCF (no cross-scale fusion) | 86.1 | 87.3 | 86.7 | 85.9 | 84.6 | 85.3 |
w/o MRGAM (no atrous conv branch) | 86.9 | 89.1 | 88.0 | 86.4 | 87.2 | 86.8 |
w/o GCT (no transformer branch) | 84.8 | 82.6 | 83.7 | 86.9 | 82.2 | 84.4 |
w/o GCT, w/ CSCF (two local streams)* | 85.3 | 84.4 | 84.8 | 86.0 | 84.1 | 85.0 |
w/o MRGAM, w/ CSCF (conv + transformer)* | 87.5 | 88.6 | 88.1 | 87.6 | 88.0 | 87.8 |
w/o GCT & w/o CSCF (MRGAM only) | 81.9 | 80.6 | 81.2 | 82.7 | 80.9 | 81.8 |
w/o MRGAM & w/o CSCF (GCT only) | 82.7 | 82.0 | 82.4 | 83.6 | 81.6 | 82.6 |
w/o GCT & w/o MRGAM (dual plain encoders) | 79.8 | 81.0 | 80.4 | 80.4 | 82.0 | 81.2 |
baseline U-Net | 77.6 | 82.7 | 80.0 | 78.3 | 83.8 | 81.0 |
