Table 4 Comparison of landmark detection precision.

From: Multimodal deep learning for cephalometric landmark detection and treatment prediction

Method

Mean detection error (mm)

Standard deviation

2 mm success rate (%)

3 mm success rate (%)

Evaluation dataset

Implementation note

DeepFuse (Ours)

1.21

0.58

92.4

97.8

Combined

Original implementation

U-Net89*

2.35

1.12

72.1

88.5

Combined

Re-implemented and evaluated on our dataset

ResNet-based90*

1.87

0.94

78.3

91.2

Combined

Re-implemented and evaluated on our dataset

DenseNet + Heatmap*

1.79

0.83

80.5

92.6

Combined

Re-implemented and evaluated on our dataset

CBCT-only

1.65

0.74

83.7

94.1

Combined

Ablation of our method

CephNet91†

1.58

0.71

85.2

94.8

CephNet

Results from original publication

You et al.92†

1.43

0.65

87.6

95.3

OrthoFace

Results from original publication

Wang et al.93†

1.39

0.62

88.9

96.2

Private dataset

Results from original publication

  1. *Methods re-implemented and evaluated on our combined dataset for direct comparison. †Results reported from original publications on different datasets; included for reference but not direct comparison.
  2. **For meaningful comparison, we re-implemented and evaluated U-Net, ResNet-based, and DenseNet approaches on our combined dataset. Performance claims for DeepFuse are based on direct comparisons with these re-implementations. Results from CephNet, You et al., and Wang et al. are provided for contextual reference but were obtained on different datasets with varying characteristics, making direct performance comparisons less valid.
Back to article page