Multiscale Feature and Attention Mechanism for Blood Vessel Segmentation in Fundus Images

Guangcen Ma; Jinzhi Zhou; Haoyang He; Saifeng Li

doi:10.3788/LOP241471

Laser & Optoelectronics Progress, Volume. 62, Issue 4, 0417001(2025)

Multiscale Feature and Attention Mechanism for Blood Vessel Segmentation in Fundus Images

Guangcen Ma^1,2、*, Jinzhi Zhou^1,2, Haoyang He^1,2, and Saifeng Li^1,2

Author Affiliations

¹School of Information Engineering, Southwest University of Science and Technology, Mianyang 621000, Sichuan , China

²Robot Technology Used for Special Environment Key Laboratory of Sichuan Province, Mianyang 621000, Sichuan , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(15)

Fig. 1. Structure of the MSF-DA-Unet network

Download full size

Fig. 2. Structure of the DRA module

Download full size

Fig. 3. Structure of the DA module

Download full size

Fig. 4. Structure of the AFF module

Download full size

Fig. 5. Preprocessing images. (a) Grayscale image; (b) standardized image; (c) CLAHE image; (d) gamma-corrected image

Download full size

Fig. 6. Image local sample blocks and corresponding gold standard blocks. (a) Integration diagram of the local sample blocks; (b) integration diagram of the corresponding gold standard blocks

Download full size

Fig. 7. Segmentation results by different algorithms. (a) Original images; (b) gold standard images; (c) MSF-DA-Unet; (d) U-Net; (e) LadderNet; (f) AttU-Net; (g) UNet++

Download full size

Fig. 8. Detailed segmentation results by different algorithms. (a) Original images; (b) gold standard images; (c) MSF-DA-Unet; (d) U-Net; (e) LadderNet; (f) AttU-Net; (g) UNet++

Download full size

Fig. 9. Segmentation results by different improved networks. (a) Original images; (b) gold standard images; (c) U-Net; (d) dilated residual module; (e) DRA module; (f) MSF-DA-Unet

Download full size

Table 1. Results comparison of the model on the DRIVE dataset under different weight factors

View table

Table 1. Results comparison of the model on the DRIVE dataset under different weight factors

Weight factor	AUC	Acc	S_e	S_p	F1 score
0.1	0.9803	0.9624	0.8061	0.9847	0.8268
0.2	0.9808	0.9625	0.8082	0.9843	0.8273
0.3	0.9812	0.9624	0.8105	0.9840	0.8275
0.4	0.9817	0.9623	0.8106	0.9840	0.8296
0.5	0.9821	0.9624	0.8109	0.9839	0.8298
0.6	0.9830	0.9625	0.8125	0.9837	0.8303
0.7	0.9834	0.9625	0.8132	0.9836	0.8306
0.8	0.9835	0.9624	0.8121	0.9837	0.8309
0.9	0.9838	0.9625	0.8108	0.9839	0.8297

Table 2. Effect comparison of different loss functions
View table
Table 2. Effect comparison of different loss functions
Loss function AUC Acc S_e S_p F1 score
Cross-entropy loss 0.9798 0.9624 0.8122 0.9832 0.8302
Focal loss 0.9841 0.9625 0.8084 0.9853 0.8293
Focal loss +Cross-entropy loss 0.9834 0.9625 0.8132 0.9836 0.8306

Table 3. Effect comparison of different attention modules
View table
Table 3. Effect comparison of different attention modules
Attention module AUC Acc S_e S_p F1 score
scSE 0.9823 0.9614 0.8070 0.9834 0.8256
DA 0.9834 0.9625 0.8132 0.9836 0.8306

Table 4. Ablation experimental results on the DRIVE dataset

View table

Table 4. Ablation experimental results on the DRIVE dataset

Method	AUC	Acc	S_e	S_p	F1 score
M1	0.9720	0.9526	0.7712	0.9756	0.8166
M2	0.9741	0.9561	0.7967	0.9819	0.8252
M3	0.9786	0.9601	0.8082	0.9812	0.8254
M4	0.9822	0.9597	0.8043	0.9823	0.8270
M5	0.9831	0.9617	0.8092	0.9840	0.8283
M6	0.9834	0.9625	0.8132	0.9836	0.8306

Table 5. Comparison of different algorithms on the DRIVE dataset

View table

Table 5. Comparison of different algorithms on the DRIVE dataset

Algorithm	AUC	Acc	S_e	S_p	F1 score
U-Net^［8］	0.9720	0.9526	0.7712	0.9756	0.8166
R2U-Net^［12］	0.9784	0.9556	0.7792	0.9813	0.8171
LadderNet^［13］	0.9795	0.9561	0.7856	0.9810	0.8202
AttU-Net^［23］	0.9722	0.9564	0.7977	0.9799	0.8236
UNet++^［24］	0.9735	0.9570	0.8041	0.9803	0.8255
DCU-net^［14］	0.9810	0.9568	0.8115	0.9780	0.8272
Shuffle-Unet^［15］	0.9782	0.9561	0.7801	0.9828	0.8155
Ref. ［16］	0.9807	0.9568	0.8054	0.9789	0.8261
ResDO-UNet^［25］	–	0.9563	0.7985	0.9791	0.8229
IterMiUnet^［26］	0.9810	0.9568	0.8053	0.9789	0.8262
MSF-DA-Unet	0.9834	0.9625	0.8132	0.9836	0.8306

Table 6. Comparison of different algorithms on the CHASE_DB1 dataset

View table

Table 6. Comparison of different algorithms on the CHASE_DB1 dataset

Algorithm	AUC	Acc	S_e	S_p	F1 score
U-Net^［8］	0.9732	0.9589	0.7707	0.9732	0.7873
R2U-Net^［12］	0.9815	0.9634	0.7756	0.9820	0.7928
LadderNet^［13］	0.9839	0.9652	0.7978	0.9818	0.8031
AttU-Net^［23］	0.9789	0.9652	0.8180	0.9799	0.8064
UNet++^［24］	0.9792	0.9657	0.7804	0.9834	0.8102
DCU-net^［14］	0.9872	0.9664	0.8075	0.9841	0.8278
Shuffle-Unet^［15］	0.9823	0.9642	0.7762	0.9825	0.7954
Ref. ［16］	0.9836	0.9635	0.8240	0.9775	0.8052
ResDO-UNet^［25］	–	0.9672	0.8020	0.9794	0.8236
IterMiUnet^［26］	0.9812	0.9591	0.8441	0.9704	0.7856
MSF-DA-Unet	0.9870	0.9696	0.8259	0.9845	0.8215

Tools

Get Citation

Copy Citation Text

Guangcen Ma, Jinzhi Zhou, Haoyang He, Saifeng Li. Multiscale Feature and Attention Mechanism for Blood Vessel Segmentation in Fundus Images[J]. Laser & Optoelectronics Progress, 2025, 62(4): 0417001

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites