Table 2 Main results on the primary TCGA-LIHC test set

From: STD-Net: a spatio-temporal decoupling network for multiphasic liver lesion segmentation and characterization

 

Segmentation

Characterization

Model

DSC (%) ↑

HD95 (mm) ↓

AUC ↑

Acc. (%) ↑

F1 ↑

CNN-based Methods

 3D U-Net

78.4

11.23

0.812

75.7

0.751

 V-Net

79.1

10.55

0.819

76.1

0.758

 nnU-Net

83.5

7.14

0.865

80.0

0.795

Transformer-based Methods

 UNETR

82.6

7.98

0.851

78.6

0.780

 Swin UNETR

84.0

6.81

0.870

81.4

0.811

 MedNeXt

84.7

6.45

0.882

82.9

0.825

 SegMamba

84.3

6.60

0.876

82.1

0.818

Foundation Model-based Methods

 Medical SAM

80.5

9.87

–

–

–

 SAM-Med3D

81.8

8.41

–

–

–

Advanced Fusion & Temporal Methods

 Cross-Attn Fusion

83.8

7.02

0.871

81.4

0.809

 I3D

80.2

10.11

0.840

78.6

0.782

 Timesformer

81.5

9.15

0.859

80.0

0.798

 ST-Adapter

82.1

8.80

0.863

80.7

0.801

 LoGoFormer

82.5

8.54

0.868

81.4

0.810

 CF-Net (2024)

84.9

6.72

0.878

82.8

0.822

 MVFusion (2024)

85.0

6.51

0.881

83.1

0.826

 CoCa-DR

85.1

6.33

0.886

83.6

0.832

Ours

STD-Net

87.2

4.12

0.924

87.1

0.868

  1. This table compares the multi-task performance for both segmentation and characterization. All models were trained and evaluated using the same three-phase (arterial, portal, delayed) inputs and identical preprocessing for fairness. For architectures not natively supporting multiphasic input (e.g., nnU-Net, MedNeXt), early channel concatenation was applied following standard practice. Best results are in bold, second-best are underlined.
Back to article page