All-model CAD prediction comparison

Training-set diagnostic only. Run 14h06, CSV-seeded weak 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.

Model	Native loss	Training AP	IoU@0.5	Dice@0.5	Top-K recall	Display coverage	Process	View
2D CNN/U-Net	2.524 → 1.084	0.298	0.143	0.250	0.308	92.2%	yes	CAD field
FNO2d	2.52 → 0.7652	0.444	0.143	0.250	0.385	92.2%	yes	CAD field
Dense 3D U-Net	2.338 → 0.5917	0.097	0.019	0.037	0.231	100.0%	yes	CAD field
Sparse 3D U-Net	2.393 → 0.6716	0.112	0.019	0.037	0.231	100.0%	no	CAD field
PointNet	2.42 → 1.486	0.076	0.009	0.017	0.154	99.7%	no	CAD field
DGCNN	2.713 → 1.298	0.038	0.065	0.122	0.077	63.0%	no	CAD field
Mesh GNN	1.85 → 1.162	0.045	0.021	0.040	0.077	100.0%	yes	CAD field
Risk GNN	1.884 → 1.023	0.253	0.036	0.070	0.308	100.0%	yes	CAD field
UV-Net	2.182 → 1.208	0.011	0.011	0.021	0.000	96.6%	no	CAD field
Tabular MLP	0.2116 → 3.465e-08	0.001	0.000	0.000	0.000	100.0%	yes	CAD field

All models mapped onto the same CAD

2D CNN/U-Net

FNO2d

Dense 3D U-Net

Sparse 3D U-Net

PointNet

DGCNN

Mesh GNN

Risk GNN

UV-Net

Tabular MLP