Table 3 Cross-dataset generalization performance on NSCLC-Radiomics and MosMedData

From: CoreFormer high fidelity pulmonary nodule segmentation with structural core priors and geodesic implicit fields

Method

NSCLC-Radiomics

MosMedData

 

Dice↑

HD↓

BF1↑

Dice↑

HD↓

BF1↑

3D U-Net

78.4

4.92

74.1

75.2

5.10

71.5

V-Net

79.0

4.78

75.0

75.6

4.85

72.2

Attention U-Net

79.5

4.60

75.8

76.1

4.73

73.0

nnU-Net

80.3

4.38

76.7

76.9

4.45

74.1

STUNet

81.1

4.02

78.2

77.5

4.13

75.2

TransUNet

79.7

4.36

76.1

76.2

4.42

73.8

UNETR

80.5

4.15

77.4

77.0

4.25

74.6

CoreFormer (Ours)

83.2

3.67

80.3

78.8

3.88

77.0

  1. All models are trained on LIDC-IDRI and directly evaluated (without fine-tuning) on the target datasets. Metrics include Dice (%), Hausdorff Distance (HD, mm), and Boundary F1 Score (BF1, %).
Back to article page