Table 1 Comparison of classification performance on INbreast and BUSI datasets

From: Anatomy-guided visual prompt tuning for cross-modal breast cancer understanding

Method

INbreast AUC

INbreast F1

BUSI AUC

BUSI F1

ResNet-5021

93.2

88.1

89.4

84.7

DenseNet-12122

94.5

89.0

90.2

85.5

TransMIL23

95.3

90.1

91.6

86.8

Swin-Transformer3

95.6

90.8

92.1

87.2

ViT-B/16 Fine-tuned2

96.1

91.3

93.5

88.0

MedCLIP6

96.4

91.5

94.0

88.5

LoRA24

96.5

91.9

94.2

88.8

Adapter-Tuning25

96.7

92.1

94.3

89.0

VPT14

96.9

92.4

94.6

89.3

CoOp26

97.0

92.6

94.8

89.4

A-VPT (ours)

97.8

93.5

95.7

90.6

  1. Metrics are reported as AUC (%)/F1 (%). Best results are in bold. Similar to adapter-tuning methods originating from natural language processing, we adopted a similar approach and applied it to visual models. Note: All PEFT methods (LoRA, Adapter, VPT, CoOp, A-VPT) utilize the same frozen ViT-B/16 backbone to ensure fair comparison.
Back to article page