Optics and Precision Engineering, Volume. 32, Issue 4, 565(2024)

Brain tumor image segmentation based on Semantic Flow Guided Sampling and Attention Mechanism

Jianli SONG1... Xiaoqi LÜ1,2,* and Yu GU1 |Show fewer author(s)
Author Affiliations
  • 1School of Information Engineering, Inner Mongolia University of Science and Technology, Baotou0400, China
  • 2School of Information Engineering, Inner Mongolia University of Technology, Hohhot010051, China
  • show less
    Figures & Tables(15)
    Overall network architecture of our DAFANet model
    Structural diagram of Multi-Fiber unit and Dilated Multi-Fiber unit
    Feature align unit
    Structure of expectation-maximization attention
    Brain tumor image and real segmentation tags with different modalities
    Comparison chart of different values of k
    Comparison of box diagram between DAFANet and DMFNet
    Visual comparison of segmentation result
    • Table 1. Datasets composition

      View table
      View in Article

      Table 1. Datasets composition

      DatasetsTrainValid
      BraTS201828566
      BraTS2019335125
    • Table 2. Comparison of effects of different models using FA

      View table
      View in Article

      Table 2. Comparison of effects of different models using FA

      MethodsDice/%Sensitivity/%Specificity/%Hausdorff95/mm
      ETWTTCETWTTCETWTTCETWTTC
      DMFNet76.6888.4780.1977.7988.3179.1499.8199.4299.733.3415.4076.138
      FA+DMFNet76.7888.5480.9877.5388.7079.4999.8299.5299.753.1345.3856.148
      3D UNet72.6887.1870.8870.1586.7265.8199.8499.4999.895.3419.1729.461
      FA+UNet74.6588.2172.6274.2887.7869.0499.8899.4799.835.4225.4309.652
      HDCNet76.9889.1581.5177.6392.2680.6199.8399.2699.754.1596.4338.009
      FA+HDCNet77.1789.5681.5578.9492.2280.7899.8199.2899.713.0795.7697.133
    • Table 3. Ablation experiment of DAFANet model

      View table
      View in Article

      Table 3. Ablation experiment of DAFANet model

      MethodsDice/%Hausdorff95/mm
      ETWTTCETWTTC
      DMFNet76.6888.4780.193.3415.4076.138
      DMFNet+EMA77.1489.6981.253.9605.1817.091
      FA+DMFNet76.7888.5480.983.1345.3856.148
      FA+DMFNet+EMA(up)76.8189.7881.613.7424.8636.954
      FA+DMFNet+EMA(down)77.0989.5981.542.6774.9596.125
      FA+DMFNet+2EMA78.1190.1082.213.0724.3606.229
    • Table 4. Ablation experiment of EMA position

      View table
      View in Article

      Table 4. Ablation experiment of EMA position

      MethodsDice/%Hausdorff95/mmParams/MFLOPs/G
      ETWTTCETWTTC
      Layer 176.7288.3578.674.3165.9726.8714.1130.51
      Layer 274.8587.6676.253.2418.4517.6284.1331.04
      Layer 378.1190.1082.213.0724.3606.2294.2330.50
    • Table 5. Comparative experiment of superparameter k

      View table
      View in Article

      Table 5. Comparative experiment of superparameter k

      kDice/%Hausdorff95/mm
      ETWTTCETWTTC
      6478.1190.1082.213.0724.3066.229
      3276.0189.3880.213.2465.0276.071
      1678.1690.0380.733.0224.6826.326
    • Table 6. Comparison of segmentation results with classical models

      View table
      View in Article

      Table 6. Comparison of segmentation results with classical models

      MethodsDice/%Hausdorff95/mmParams/MFLOPs/G
      ETWTTCETWTTC
      TransBTS1378.9390.0081.943.7365.6446.04932.99333.00
      3D UNet1072.6887.1870.885.3419.1729.46113.12176.61
      3D ESPNet2577.3589.3081.834.1906.7407.9803.6376.51
      Attention UNet1275.9688.8177.205.0207.7568.25834.9051.30
      DMFNet2076.6888.4780.193.3415.4076.1383.8827.04
      HDCNet2176.9889.1581.514.1596.4338.0090.2924.00
      DAFANet78.1190.1082.213.0724.3606.2294.2330.50
    • Table 7. Comparison of Dice coefficient and Hausdorff95 distance with other models under different datasets

      View table
      View in Article

      Table 7. Comparison of Dice coefficient and Hausdorff95 distance with other models under different datasets

      DatasetMethodsDice/%Hausdorff95/mm
      ETWTTCETWTTC
      BraTS2019Nuechterlein2577.3589.3081.834.196.747.98
      Chen2076.6888.4780.193.345.416.14
      Zhang2670.9087.0070.90N/AN/AN/A
      Sheng2772.4087.5078.805.799.3511.47
      Akbar2874.2088.4880.986.6710.2510.83
      Liu2977.9189.9483.894.035.456.56
      Chang3078.2089.0081.203.828.537.43
      Ours78.1190.1082.213.074.366.23
      BraTS2018Nuechterlein2573.7088.3081.405.305.467.85
      Chen2080.1190.6184.543.064.666.44
      Zhang2677.2087.2080.805.6912.609.62
      Sheng2777.1087.6081.103.228.4210.56
      Akbar2877.7189.5979.773.909.138.67
      Liu2980.4189.8585.442.244.055.76
      Chang3079.5090.0083.902.926.515.71
      Ours80.4490.0784.572.754.705.49
    Tools

    Get Citation

    Copy Citation Text

    Jianli SONG, Xiaoqi LÜ, Yu GU. Brain tumor image segmentation based on Semantic Flow Guided Sampling and Attention Mechanism[J]. Optics and Precision Engineering, 2024, 32(4): 565

    Download Citation

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

    Category:

    Received: Apr. 5, 2023

    Accepted: --

    Published Online: Apr. 2, 2024

    The Author Email: LÜ Xiaoqi (lxiaoqi@imut.edu.cn)

    DOI:10.37188/OPE.20243204.0565

    Topics