← model structures and inputs · JSON report · mapped fields NPZ

All models mapped onto the same CAD

Training-set diagnostic only. Run 18h01, one weak point-dilated target and unvalidated registration. AP/IoU/Dice below measure fit to this same run, not generalization. Native resolution and transfer back to CAD differ by model. Gray CAD nodes are more than 10 mm from direct mapped support.

All model risk fields on a common CAD projection
ModelNative lossTraining APIoU@0.5Dice@0.5Top-21 recallDisplay coverageProcessView
2D CNN/U-Net1.784 → 0.48210.6220.3330.5000.71498.7%yesCAD field
FNO2d1.832 → 0.49890.6200.2960.4570.71498.7%yesCAD field
Dense 3D U-Net2.273 → 0.17180.3580.0480.0930.524100.0%yesCAD field
Sparse 3D U-Net2.323 → 0.5010.3590.0470.0910.524100.0%noCAD field
PointNet1.72 → 1.1350.0930.1140.2040.000100.0%noCAD field
DGCNN1.68 → 0.01450.0020.0000.0000.00073.6%noCAD field
Mesh GNN2.039 → 1.0010.7580.1470.2560.714100.0%yesCAD field
Risk GNN1.962 → 0.560.7640.3090.4720.714100.0%yesCAD field
UV-Net2.322 → 1.2510.0760.0430.0820.00097.9%noCAD field
Tabular MLP0.2108 → 9.671e-080.0020.0000.0000.000100.0%yesCAD field

2D CNN/U-Net

2D CNN/U-Net risk field mapped onto CAD

FNO2d

FNO2d risk field mapped onto CAD

Dense 3D U-Net

Dense 3D U-Net risk field mapped onto CAD

Sparse 3D U-Net

Sparse 3D U-Net risk field mapped onto CAD

PointNet

PointNet risk field mapped onto CAD

DGCNN

DGCNN risk field mapped onto CAD

Mesh GNN

Mesh GNN risk field mapped onto CAD

Risk GNN

Risk GNN risk field mapped onto CAD

UV-Net

UV-Net risk field mapped onto CAD

Tabular MLP

Tabular MLP risk field mapped onto CAD