Acta Optica Sinica, Volume. 43, Issue 14, 1418001(2023)

Retinal Vessel Segmentation via Self-Adaptive Compensation Network

Lin Zhang1, Chuang Wu1, Xinyu Fan1, Chaoju Gong1,2, Suyan Li1,2, and Hui Liu1、*
Author Affiliations
  • 1School of Information and Control Engineering, China University of Mining and Technology, Xuzhou 221116, Jiangsu, China
  • 2Department of Ophthalmology, The First People's Hospital of Xuzhou, Xuzhou 221116, Jiangsu, China
  • show less
    Figures & Tables(13)
    Overall structure of SACom model
    Deformable convolution
    Adaptive multi-scale aligned context module
    Collaborative compensation module
    Feature layer averaging adaptive fusion
    Visualization results on different datasets. (a) Original images; (b) ground truth by expert 1; (c) ground truth by expert 2; (d) segmentation results of U-Net; (e) segmentation results of SACom
    Detail comparison of DRIVE dataset segmentation results. (a) Original images; (b) local images; (c) local images of ground truth 1; (d) local images of ground truth 2; (e) local images of SACom segmentation results
    Visualization of AUC for SACom and other advanced methods
    • Table 1. Performance comparison under different weights

      View table

      Table 1. Performance comparison under different weights

      WeightAccSeSpF1AUC
      1.00. 97030.80020.98660.82520.9878
      1.50.97000.81560.98480.82630.9875
      2.00.96950.84030.98190.82850.9880
      2.50.96810.85490.97900.82430.9874
      3.00.96690.87230.97590.82180.9878
    • Table 2. Comparison of different supervision and fusion strategies

      View table

      Table 2. Comparison of different supervision and fusion strategies

      StrategyAccSeSpF1AUC
      10.96880.83740.98140.82460.9871
      20.96840.83690.98100.82270.9868
      30.96920.83970.98160.82690.9875
      40.96950.84030.98190.82850.9880
    • Table 3. Comparison of ablation experimental results

      View table

      Table 3. Comparison of ablation experimental results

      ModelAccSeSpF1AUCParams /M
      U-Net640.95510.78410.98000.81670.975817.26
      U-Net32(Baseline)0.95540.78490.98020.81750.97614.32
      Baseline+DC1230.96420.78830.98100.79390.97894.20
      Baeline+DC2340.96470.79140.98140.79720.98023.79
      Baeline+DC3450.96120.78690.97790.78020.97653.87
      Baeline+DC234+AMAC0.96670.82770.98010.81340.98494.17
      Baeline+DC234+AMAC+CCB0.96950.84030.98190.82850.98806.01
    • Table 4. Performance comparison of SACom with different expert annotations on DRIVE test set

      View table

      Table 4. Performance comparison of SACom with different expert annotations on DRIVE test set

      LabelAccSeSpF1AUC
      10.96950.84030.98190.82850.9800
      20.97370.87130.98320.84840.9910
    • Table 5. Segmentation performance comparison of different methods

      View table

      Table 5. Segmentation performance comparison of different methods

      DatasetMethodAccSeSpF1AUC
      DRIVEU-Net0.95540.78490.98020.81750.9761
      DUNet0.95660.79630.98000.82370.9802
      AA-UNet0.95580.79410.97980.82160.9847
      CSU-Net0.95650.80710.97820.82510.9801
      CMNet0.96640.80690.98180.80640.9840
      SACom0.96950.84030.98190.82850.9880
      CHASE_DB1U-Net0.95760.73040.97170.77930.9782
      DUNet0.96100.81550.97520.78830.9804
      AA-UNet0.96080.81760.97040.78920.9865
      CSU-Net0.97060.84270.98360.81050.9824
      CMNet0.97420.84610.98360.80970.9879
      SACom0.97630.87480.98310.82290.9917
      STAREU-Net0.96360.76380.98670.81350.9789
      DUNet0.96410.75950.98780.81430.9832
      AA-UNet0.96400.75980.98780.81420.9824
      CSU-Net0.97020.84320.98450.85160.9825
      SACom0.97530.85060.98560.83940.9919
    Tools

    Get Citation

    Copy Citation Text

    Lin Zhang, Chuang Wu, Xinyu Fan, Chaoju Gong, Suyan Li, Hui Liu. Retinal Vessel Segmentation via Self-Adaptive Compensation Network[J]. Acta Optica Sinica, 2023, 43(14): 1418001

    Download Citation

    EndNote(RIS)BibTexPlain Text
    Save article for my favorites
    Paper Information

    Category: Microscopy

    Received: Feb. 27, 2023

    Accepted: Apr. 6, 2023

    Published Online: Jul. 13, 2023

    The Author Email: Hui Liu (hui.liu@cumt.edu.cn)

    DOI:10.3788/AOS230599

    Topics