SAU-Net: Multiorgan Image Segmentation Method Improved Using Squeeze Attention Mechanism

Guogang Cao; Hongdong Mao; Shu Zhang; Ying Chen; Cuixia Dai

doi:10.3788/LOP202259.0417003

Journals >Laser & Optoelectronics Progress >Volume 59 >Issue 4 >Page 0417003 > Article

Laser & Optoelectronics Progress
Vol. 59, Issue 4, 0417003 (2022)

SAU-Net: Multiorgan Image Segmentation Method Improved Using Squeeze Attention Mechanism

Guogang Cao^1、*, Hongdong Mao¹, Shu Zhang¹, Ying Chen¹, and Cuixia Dai²

Author Affiliations

¹School of Computer Science and Information Engineering, Shanghai Institute of Technology, Shanghai 201418, China

²College of Sciences, Shanghai Institute of Technology, Shanghai 201418, China

show less

DOI: 10.3788/LOP202259.0417003 Cite this Article Set citation alerts

Guogang Cao, Hongdong Mao, Shu Zhang, Ying Chen, Cuixia Dai. SAU-Net: Multiorgan Image Segmentation Method Improved Using Squeeze Attention Mechanism[J]. Laser & Optoelectronics Progress, 2022, 59(4): 0417003 Copy Citation Text

show less

Fig. 1. Architecture of SAU-Net

Download full size

Fig. 2. Architecture of squeeze attention module

Download full size

Fig. 3. Number of voxels in each organ of head and neck CT images

Download full size

Fig. 4. Number of annotated organs in training dataset

Download full size

Fig. 5. Comparison of loss value curves of different models

Download full size

Fig. 6. Comparison of visualization for SAU-Net segmentation results. (a) The cross-sectional view of prediction; (b) the cross-sectional view of ground truth; (c) the cross-sectional view of overlap between prediction and ground truth; (d) the 3D view of overlap between prediction and ground truth

Download full size

Type	Source	Number
Training dataset	MICCAI 2015 training dataset	38
	Head and Neck Cetuximab collections	46
	Public dataset of the Quebec Institute of Canada	177
Test dataset	MICCAI 2015 test dataset	10

Table 1. Source and distribution of dataset

Organ	Baseline	Baseline+Res	Baseline+SE	SAU-Net
Brainstem	0.773	0.834	0.857	0.884
Chiasm	0.484	0.495	0.522	0.554
Mandible	0.830	0.858	0.917	0.932
Optic.L	0.662	0.692	0.712	0.738
Optic.R	0.628	0.683	0.709	0.729
Paro.L	0.787	0.818	0.863	0.887
Paro.R	0.768	0.824	0.859	0.881
Subm.L	0.723	0.743	0.798	0.812
Subm.R	0.732	0.738	0.805	0.820
Average	0.707	0.743	0.782	0.804

Table 2. Comparison of DSC with different models

Organ	Baseline		Baseline+Res		Baseline+SE		SAU-Net
Organ	FPR	FNR	FPR	FNR	FPR	FNR	FPR	FNR
Brainstem	0.117	0.138	0.110	0.124	0.072	0.101	0.063	0.091
Chiasm	0.105	0.461	0.091	0.457	0.081	0.347	0.074	0.292
Mandible	0.127	0.119	0.127	0.093	0.097	0.058	0.106	0.047
Optic.L	0.118	0.428	0.106	0.415	0.092	0.327	0.034	0.295
Optic.R	0.121	0.437	0.126	0.397	0.114	0.294	0.108	0.313
Paro.L	0.137	0.154	0.120	0.139	0.088	0.131	0.076	0.092
Paro.R	0.141	0.161	0.131	0.148	0.112	0.129	0.090	0.123
Subm.L	0.115	0.146	0.109	0.139	0.092	0.122	0.084	0.117
Subm.R	0.129	0.152	0.116	0.142	0.098	0.118	0.101	0.071
Average	0.123	0.244	0.115	0.228	0.094	0.181	0.082	0.160

Table 3. Comparison of FPR and FNR with different models

Organs	Ref.［13］	Ref.［14］	Ref.［2］	AnatomyNet^［17］	FocusNet^［19］	FocusNetv2^［21］	SAU-Net
Brainstem	0.880	-	0.903	0.867	0.875	0.882	0.884
Chiasm	0.557	0.580	-	0.532	0.596	0.713	0.554
Mandible	0.930	-	0.944	0.925	0.935	0.947	0.932
Optic.L	0.634	0.720	-	0.721	0.735	0.790	0.738
Optic.R	0.639	0.700	-	0.706	0.744	0.817	0.729
Paro.L	0.827	-	0.823	0.881	0.863	0.898	0.887
Paro.R	0.814	-	0.829	0.874	0.879	0.881	0.881
Subm.L	0.723	-	-	0.814	0.798	0.840	0.812
Subm.R	0.723	-	-	0.813	0.801	0.838	0.820
Average	0.749	-	-	0.793	0.803	0.845	0.804

Table 4. Comparison of DSC score of different methods

Organs	Ref.［13］	Ref.［14］	AnatomyNet^［17］	FocusNet^［19］	FocusNetv2^［21］	SAU-Net
Brainstem	4.59	-	6.42 $\pm$ 2.38	2.14 $\pm$ 0.6	2.32 $\pm$ 0.7	2.02 $\pm$ 0.8
Chiasm	2.78	2.81 $\pm$ 1.6	5.76 $\pm$ 2.49	3.16 $\pm$ 1.3	2.25 $\pm$ 0.8	2.65 $\pm$ 1.3
Mandible	1.97	-	6.28 $\pm$ 2.21	1.18 $\pm$ 0.3	1.08 $\pm$ 0.4	2.12 $\pm$ 0.7
Optic.L	2.26	2.33 $\pm$ 0.8	4.85 $\pm$ 2.32	3.76 $\pm$ 2.9	1.92 $\pm$ 0.8	2.14 $\pm$ 1.8
Optic.R	3.15	2.13 $\pm$ 1.0	4.77 $\pm$ 4.27	2.65 $\pm$ 1.5	2.17 $\pm$ 0.7	2.32 $\pm$ 1.4
Paro.L	5.11	-	9.31 $\pm$ 3.32	2.52 $\pm$ 1.0	1.81 $\pm$ 0.4	2.54 $\pm$ 1.1
Paro.R	6.13	-	10.08 $\pm$ 5.09	2.07 $\pm$ 0.8	2.43 $\pm$ 2.0	2.32 $\pm$ 1.7
Subm.L	5.35	-	7.01 $\pm$ 4.44	2.67 $\pm$ 1.3	2.84 $\pm$ 1.2	2.27 $\pm$ 1.6
Subm.R	5.44	-	6.02 $\pm$ 1.08	3.41 $\pm$ 1.4	2.74 $\pm$ 1.2	2.92 $\pm$ 1.8
Average	4.14	-	6.30	2.62	2.17	2.37

Table 5. Comparison of 95HD score of different methods

Models	Metrics
Models	Parameters /million	Time /s
AnatomyNet	0.73	0.68
FocusNetv2	2.02	1.88
SAU-Net	0.79	0.51

Table 6. Comparison of parameters and inference time of different methods

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information