Acta Optica Sinica, Volume. 40, Issue 6, 0610002(2020)
Retinal Vascular Image Segmentation Based on Improved HED Network
Figures & Tables(13)
Fig. 2. Schematic diagram of the sampling position of a 5×5 ordinary convolution and a 5×5 deformable convolution. (a) Image after preprocessing; (b) ordinary convolution; (c) deformable convolution
Fig. 3. Schematic of dilated convolution. (a) Dilation rate is 1; (b) dilation rate is 2; (c) dilation rate is 3
Fig. 5. Typical image after preprocessing. (a) Original image; (b) merged image between a red channel and a green channel; (c) image after CLAHE operation; (d) image after Gamma correction
Fig. 6. Segmentation results on DRIVE. (a) Original images; (b) ground truth; (c) segmentation result images
Fig. 7. Segmentation results on STARE. (a) Original images; (b) ground truth; (c) segmentation result images
Fig. 8. Segmentation results on lesion images. (a) Original images; (b) ground truth; (c) segmentation result images by proposed method; (d) segmentation result images in Ref. [24]
Fig. 9. Local maps segmented by different models. (a) Original image; (b) standard local map; (c)-(f) local maps segmented by model 1 to model 4, respectively; (g) local map segmented by the proposed method
Table 1. Segmentation results on DRIVE and STARE
View tableTable 1. Segmentation results on DRIVE and STARE
Database Method Sn /% Sp /% Acc /% AUC /% DRIVE 2nd Human observer 77.96 97.17 94.64 - Proposed method 81.75 97.67 95.44 98.33 STARE 2nd Human observer 89.52 93.84 93.49 - Proposed method 80.68 98.38 96.56 98.12
Table 2. Performance comparison of the proposed method with state-of-the-art methods on the DRIVE
View tableTable 2. Performance comparison of the proposed method with state-of-the-art methods on the DRIVE
Type Method Year Sn /% Sp /% Acc /% AUC /% Unsupervised method Zhang et al.[ 3 ]2010 71.77 97.24 93.82 — Zhao et al.[ 5 ]2015 74.20 98.20 95.40 86.20 Dash et al.[ 4 ]2019 70.30 98.50 95.10 — Supervised method Liskowski et al.[ 17 ]2016 75.69 98.16 95.27 97.38 Lu et al.[ 9 ]2018 78.12 98.14 95.59 97.90 Wang et al.[ 19 ]2019 76.48 98.17 95.41 — Yan et al.[ 22 ]2019 76.31 98.20 95.38 97.50 Proposed method 2019 81.75 97.67 95.44 98.33
Table 3. Performance comparison of the proposed method with state-of-the-art methods on the STARE
View tableTable 3. Performance comparison of the proposed method with state-of-the-art methods on the STARE
Type Method Year Sn /% Sp /% Acc /% AUC /% Unsupervised method Zhao et al.[ 5 ]2015 78.00 97.80 95.60 87.40 Khan et al.[ 23 ]2017 73.59 97.08 95.02 — Supervised method Liskowski et al.[ 17 ]2016 70.27 98.28 95.45 96.71 Fu et al.[ 18 ]2016 74.12 — 95.85 — Wang et al.[ 19 ]2019 75.23 98.85 96.40 — Yan et al.[ 22 ]2019 77.35 98.57 96.38 98.33 Proposed method 2019 80.68 98.38 96.56 98.12
Table 4. Performance results of different HED models on the DRIVE and STARE
View tableTable 4. Performance results of different HED models on the DRIVE and STARE
Model DRIVE STARE Se /% Sp /% Acc /% AUC /% Se /% Sp /% Acc /% AUC /% Model1 69.70 98.96 94.89 97.62 63.60 98.99 94.92 97.40 Model2 79.55 97.82 95.23 98.13 73.61 98.36 95.45 97.78 Model3 77.08 98.21 95.27 97.79 72.74 98.61 96.02 97.65 Model4 78.48 98.23 95.45 97.93 75.13 99.05 96.35 98.81 Our method 81.75 97.67 95.44 98.33 80.68 98.38 96.56 98.12
Tools
Get Citation
Copy Citation Text
Sai Zhang, Yanping Li. Retinal Vascular Image Segmentation Based on Improved HED Network[J]. Acta Optica Sinica, 2020, 40(6): 0610002
Paper Information
Category: Image Processing
Received: Nov. 1, 2019
Accepted: Nov. 29, 2019
Published Online: Mar. 6, 2020
The Author Email: Li Yanping (z17835422201@163.com)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20