Multi-Scale Residual U-Net Fundus Blood Vessel Segmentation Based on Attention Mechanism

Feng Zhao; Beibei Zhong; Hanqiang Liu

doi:10.3788/LOP202259.1810002

Laser & Optoelectronics Progress, Volume. 59, Issue 18, 1810002(2022)

Multi-Scale Residual U-Net Fundus Blood Vessel Segmentation Based on Attention Mechanism

Feng Zhao¹, Beibei Zhong^1、*, and Hanqiang Liu²

¹School of Communication and Information Engineering & School of Artificial Intelligence, Xi’an University of Posts and Telecommunications, Xi’an 710121, Shaanxi , China

²School of Computer Science, Shaanxi Normal University, Xi’an , Shaanxi 710119, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(16)

Fig. 1. Network structure comparison. (a) U-Net; (b) multi-scale residual U-shaped network based on attention mechanism

Download full size

Fig. 2. Improved residual block structure

Download full size

Fig. 3. Multi-scale convolution module

Download full size

Fig. 4. Parallel dilated convolution module

Download full size

Fig. 5. Multi-scale attention module

Download full size

Fig. 6. Hybrid attention module

Download full size

Fig. 7. Image preprocessing. (a) Original image of DRIVE dataset; (b) pre-processed image

Download full size

Fig. 8. Retinal vessel segmentation results of different algorithms. (a) Original images; (b) ground truth;(c) proposed algorithm; (d) Residual U-Net^[12]; (e) Recurrent U-Net^[12]; (f) R2U-Net^[12]; (g) algorithm in reference [28]

Download full size

Fig. 9. Detail comparison of segmentation results. (a) Original image; (b) details of original images; (c) details of ground truth; (d) details of proposed algorithm; (e) details of Residual U-Net^[12]; (f) details of Recurrent U-Net^[12]; (g) details of R2U-Net^[12]; (h) details of algorithm reference [28]

Download full size

Fig. 10. Verification of role of a single module. (a) Original images; (b) Ground truth; (c) M1; (d) M2; (e) M3; (f) M4

Download full size

Table 1. Formula of evaluation index
View table
Table 1. Formula of evaluation index
Index Formula
SEN $R_{S E N} = \frac{N_{T P}}{N_{T P} + N_{F N}}$
SPE $R_{S P E} = \frac{N_{T N}}{N_{T N} + N_{F P}}$
F1-score $S_{F 1 - s c o r e} = \frac{2 N_{T P}}{2 N_{T P} + N_{F N} + N_{F P}}$
ACC $R_{A C C} = \frac{N_{T N} + N_{T P}}{N_{T P} + N_{F P} + N_{T N} + N_{F N}}$

Table 2. Verification experiment of role of a single module
View table
Table 2. Verification experiment of role of a single module
Method SEN SPE F1 ACC AUC
M1 0.7736 0.9822 0.7901 0.9640 0.9754
M2 0.7638 0.9891 0.8141 0.9694 0.9849
M3 0.7830 0.9870 0.8158 0.9690 0.9845
M4 0.7757 0.9873 0.8114 0.9684 0.9829

Table 3. Multi-module cumulative effect verification experiment

View table

Table 3. Multi-module cumulative effect verification experiment

Method	SEN	SPE	F1	ACC	AUC
N1	0.7736	0.9822	0.7901	0.9640	0.9754
N2	0.7638	0.9891	0.8141	0.9694	0.9849
N3	0.8059	0.9861	0.8265	0.9703	0.9865
N4	0.8188	0.9863	0.8285	0.9704	0.9869
N5	0.8017	0.9868	0.8268	0.9706	0.9872
Proposed method	0.8267	0.9851	0.8308	0.9707	0.9876

Table 4. Validation experiment of hybrid attention module
View table
Table 4. Validation experiment of hybrid attention module
Model SEN SPE F1 ACC AUC
Spatial-channel attention 0.8245 0.9844 0.8299 0.9704 0.9871
Channel-spatial attention 0.8267 0.9851 0.8308 0.9707 0.9876

Table 5. DRIVE dataset fundus blood vessel segmentation results

View table

Table 5. DRIVE dataset fundus blood vessel segmentation results

Type	Method	Year	SEN	SPE	F1	ACC	AUC
Unsupervised method	Reference ［6］	2010	0.7120	0.9724		0.9382
	Reference ［5］	2014	0.6280	0.9840		0.9380
	Reference ［7］	2019	0.7030	0.9850		0.9510
Supervised method	Residual U-Net^［12］	2018	0.7726	0.9820	0.8149	0.9553	0.9779
	Recurrent U-Net^［12］	2018	0.7751	0.9816	0.8155	0.9556	0.9782
	R2U-Net^［12］	2018	0.7792	0.9813	0.8171	0.9556	0.9784
	Reference ［28］	2018	0.7730	0.9823	0.8148	0.9676	0.9725
	Reference ［13］	2018	0.7844	0.9819		0.9567	0.9807
	Reference ［14］	2019	0.8038	0.9802		0.9578	0.9821
	Reference ［15］	2019	0.8100	0.9848		0.9692	0.9856
	Reference ［16］	2020	0.8062	0.9769		0.9547	0.9739
	Reference ［32］	2020	0.7651	0.9818		0.9547	0.9750
	Proposed method	2021	0.8267	0.9851	0.8308	0.9707	0.9876

Table 6. CHASE DB1 dataset fundus blood vessel segmentation results

View table

Table 6. CHASE DB1 dataset fundus blood vessel segmentation results

Type	Method	Year	SEN	SPE	F1	ACC	AUC
Unsupervised method	Reference［33］	2015	0.7201	0.9824		0.9530	0.9532
Unsupervised method	Reference［34］	2018	0.7555	0.9807		0.9521
Supervised method	Residual U-Net^［12］	2018	0.7726	0.9820	0.7800	0.9553	0.9779
	Recurrent U-Net^［12］	2018	0.7459	0.9836	0.7810	0.9622	0.9803
	R2U-Net^［12］	2018	0.7756	0.9820	0.7928	0.9634	0.9815
	Reference［28］	2018	0.7820	0.9850	0.8012	0.9680	0.9819
	Reference［13］	2018	0.7538	0.9847		0.9637	0.9825
	Reference［14］	2019	0.8132	0.9814		0.9661	0.9860
	Reference［15］	2019	0.8186	0.9848		0.9743	0.9863
	Reference［16］	2020	0.8135	0.9762		0.9617	0.9782
	Reference［35］	2020	0.8477	0.9825	0.8652	0.9643	0.9448
	Proposed method	2021	0.8520	0.9850	0.8201	0.9765	0.9911

Tools

Get Citation

Copy Citation Text

Feng Zhao, Beibei Zhong, Hanqiang Liu. Multi-Scale Residual U-Net Fundus Blood Vessel Segmentation Based on Attention Mechanism[J]. Laser & Optoelectronics Progress, 2022, 59(18): 1810002

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites