Author Affiliations
1College of Communication and Infornation Technology, Xi'an University of Posts and Telecommunications, Xi'an, Shaanxi 710121, China2Key Laboratory of Ministry of Public Security, Electronic Information Field Inspection and Application Technology, Xi'an, Shaanxi 710121, Chinashow less
Fig. 1. Improved U-Net retinal vessel segmentation algorithm model
Fig. 2. Classic Inception structure
Fig. 3. Inception module
Fig. 4. Schematic diagram of hole convolution under different expansion rates r. (a) r=1;(b) r=2;(c) r=4
Fig. 5. Schematic diagram of cascaded dilated convolution module
Fig. 6. Internal structure of attention mechanism
Fig. 7. DRIVE dataset (from left to right are the original color fundus image, two expert manual segmentation images, and binary mask image)
Fig. 8. Retina image preprocessing. (a) Original image of the DRIVE dataset; (b) pre-processed image
Fig. 9. Local blocky information map of retinal blood vessels. (a) Block information of the DRIVE dataset; (b) standard block information
Fig. 10. Segmentation of experimental results. (a) Original image preprocessing map; (b) image segmentation standard map; (c) experimental result segmentation map
Fig. 11. Partial blood vessel region segmentation diagram. (a) Original color fundus retinal images; (b) locally fundus retinal images; (c) local standard retinal segmentation images; (d) local retinal segmentation result images
Fig. 12. Comparison of evaluation indexes of different algorithms
Algorithm | αSen | αSpe | αAcc | AUC |
---|
Manual segmentation method by 2th observer | 0.7796 | 0.9717 | 0.9464 | 0.9466 | Method in Ref. [8] | 0.742 | 0.982 | 0.954 | 0.862 | Method in Ref. [14] | 0.7648 | 0.9817 | 0.9541 | - | Method in Ref. [15] | 0.7763 | 0.9768 | 0.9495 | 0.972 | Method in Ref. [20] | 0.7631 | 0.982 | 0.9538 | 0.975 | Method in Ref. [21] | 0.8053 | 0.9767 | 0.9546 | 0.9771 | Method in Ref. [30] | 0.7655 | 0.9704 | 0.9442 | 0.9614 | Method in Ref. [31] | 0.8173 | 0.9733 | 0.9767 | 0.9475 | Our algorithm | 0.8274 | 0.9871 | 0.9643 | 0.9869 |
|
Table 1. Comparison of test results of different algorithms on the DRIVE dataset
Algorithm | αSen | αSpe | αAcc | AUC |
---|
a | 0.7659 | 0.9804 | 0.9528 | 0.9765 | b | 0.8014 | 0.9853 | 0.9562 | 0.9812 | c | 0.7856 | 0.9832 | 0.9546 | 0.9784 | Our algorithm | 0.8274 | 0.9871 | 0.9643 | 0.9869 |
|
Table 2. Comparison of test results of different network structures on the DRIVE dataset