Pulmonary Embolism Risk Stratification on CTPA and Medical Records dataset (val)
89.2F1 ScoreCross-attention in Graph Transformers
Evaluation results
| Method | Links | |
|---|---|---|
| Cross-attention in Graph TransformersType=GT, Fusion method=Cross-attention, Best backbone=GPS, # params=932K, Global features=✓, Vascular graph=✓2026.06 | 89.2 | |
| GPS (Early fusion)Type=GT, Fusion method=Early fusion, Best backbone=GPS, # params=35.3K, Global features=✓, Vascular graph=✓2026.06 | 88.3 | |
| GCN+VN (Early fusion)Type=MPNN, Fusion method=Early fusion, Best backbone=GCN+VN, # params=4.3K, Global features=✓, Vascular graph=✓2026.06 | 88.1 | |
| GIN+VN (Late fusion)Type=MPNN, Fusion method=Late fusion, Best backbone=GIN+VN, # params=89K, Global features=✓, Vascular graph=✓2026.06 | 88.1 | |
| Virtual Node from Global FeaturesType=MPNN, Fusion method=Virtual node, Best backbone=GIN+VN, # params=8.4K, Global features=✓, Vascular graph=✓2026.06 | 87.6 | |
| GPS (Late fusion)Type=GT, Fusion method=Late fusion, Best backbone=GPS, # params=97.4K, Global features=✓, Vascular graph=✓2026.06 | 87.5 | |
| TabPFNType=Tabular, Fusion method=-, Best backbone=TabPFN, # params=7.2M, Global features=✓, Vascular graph=-2026.06 | 86.3 | |
| GPSType=GT, Fusion method=-, Best backbone=GPS, # params=23K, Global features=-, Vascular graph=✓2026.06 | 65.6 | |
| GCN+VNType=MPNN, Fusion method=-, Best backbone=GCN+VN, # params=22.3K, Global features=-, Vascular graph=✓2026.06 | 65.5 |