Table 4 Performance summary across datasets (higher Dice is better, lower HD95 is better)
From: HemaContour: explicit parametric contour learning for robust ICH segmentation on non-contrast CT
ID | Model | INSTANCE | PhysioNet CT–ICH | ||
|---|---|---|---|---|---|
Dice (%) ↑ | HD95 (mm) ↓ | Dice (%) ↑ | HD95 (mm) ↓ | ||
B1 | nnU-Net (2D) | 83.1 | 9.6 | 79.8 | 11.2 |
B2 | nnU-Net (3D) | 84.6 | 8.7 | 81.0 | 10.4 |
B3 | 3D U-Net | 82.4 | 10.1 | 78.6 | 11.8 |
B4 | V-Net | 81.2 | 10.8 | 77.5 | 12.2 |
B5 | Attention U-Net (2D) | 83.8 | 9.3 | 80.2 | 10.9 |
B6 | UNet++ (2D) | 83.5 | 9.4 | 79.9 | 11.0 |
B7 | UNETR (3D) | 84.0 | 9.0 | 80.6 | 10.7 |
B8 | Swin-UNETR (3D) | 85.0 | 8.5 | 81.8 | 9.9 |
B9 | TransUNet (2D) | 83.0 | 9.7 | 79.4 | 11.3 |
B10 | Deep Snake (2D) | 82.7 | 8.9 | 79.1 | 10.1 |
Ours | HemaContour | 87.2 | 7.3 | 84.3 | 8.5 |
