Efficient Multi-Scale Attention Decoding Method for Thyroid Nodule Segmentation Based on SwinTransCAD

Yunpeng Wang; Jincao Yao; Dong Xu; Xiang Hao

doi:10.3788/LOP242250

Laser & Optoelectronics Progress, Volume. 62, Issue 12, 1237004(2025)

Efficient Multi-Scale Attention Decoding Method for Thyroid Nodule Segmentation Based on SwinTransCAD

Yunpeng Wang¹, Jincao Yao^2,3、***, Dong Xu^2,3、**, and Xiang Hao^1、*

Author Affiliations

¹College of Optical Science and Engineering, Zhejiang University, Hangzhou 310027, Zhejiang , China

²Department of Ultrasound, Zhejiang Cancer Hospital, Hangzhou 310022, Zhejiang , China

³Hangzhou Institute of Medicine, Chinese Academy of Sciences, Hangzhou 310000, Zhejiang , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(6)

Fig. 1. SwinTransCAD model structure framework diagram. (a) Transformer encoder; (b) EM encoder; (c) SWT encoder; (d) EUCB; (e) MSCAM; (f) MSCB; (g) LGAG; (h) CAB; (i) SAB

Download full size

Fig. 2. SwinTransCAD segmentation model results display

Download full size

Fig. 3. Comparison of thyroid nodule segmentation performance of various models on the IRB-44 dataset. (a) Thyroid ultrasound image; (b) Swin Transformer model segmentation result; (c) Swin-UNet model segmentation result; (d) radiologist annotated ground truth; (e) SwinTransCAD model segmentation result; (f) U-Net model segmentation result

Download full size

Fig. 4. ROC curves of five models

Download full size

Table 1. Comparison of segmentation performance of multiple models across different datasets

View table

Table 1. Comparison of segmentation performance of multiple models across different datasets

Method	Dataset	Sensitivity	Accuracy	DC	AUC
U-Net	IRB-44	0.883	0.947	0.863	0.948
	TN3K	0.796	0.897	0.788	0.892
	DDTI	0.754	0.883	0.745	0.881
DeepLabV3+	IRB-44	0.853	0.939	0.878	0.946
	TN3K	0.848	0.921	0.833	0.923
	DDTI	0.820	0.917	0.814	0.921
Swin Transformer	IRB-44	0.886	0.951	0.883	0.952
	TN3K	0.833	0.913	0.828	0.909
	DDTI	0.824	0.899	0.810	0.897
SwinTransCAD	IRB-44	0.902	0.962	0.905	0.961
	TN3K	0.872	0.939	0.868	0.941
	DDTI	0.881	0.952	0.878	0.954
TransUNet	IRB-44	0.883	0.933	0.881	0.933
Swin-UNet	IRB-44	0.879	0.912	0.874	0.929
UNeXt	IRB-44	0.881	0.915	0.871	0.921

Table 2. Ablation study results
View table
Table 2. Ablation study results
Component Average DC /%
None 88.33±0.3
EMCAD 88.48±0.2
EMCAD+LGAG 88.57±0.2
EMCAD+MSCAM 89.03±0.3
EMCAD+LGAG+MSCAM 89.62±0.3

Tools

Get Citation

Copy Citation Text

Yunpeng Wang, Jincao Yao, Dong Xu, Xiang Hao. Efficient Multi-Scale Attention Decoding Method for Thyroid Nodule Segmentation Based on SwinTransCAD[J]. Laser & Optoelectronics Progress, 2025, 62(12): 1237004

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites