Chinese Journal of Lasers, Volume. 51, Issue 21, 2107110(2024)

Full‐Automatic Brain Tumor Segmentation Based on Multimodal Feature Recombination and Scale Cross Attention Mechanism

Hengyi Tian, Yu Wang*, and Hongbing Xiao**
Author Affiliations
  • School of Computer and Artificial Intelligence, Beijing Technology and Business University, Beijing 100048, China
  • show less
    Figures & Tables(12)
    Framework of overall network
    Structure schematic of MR module
    Structure schematic of SC-UNet
    Structure schematics. (a) Residual convolution module; (b) convolution module
    Structure schematic of Transformer module.
    Four modal MRI images and physician-labeled brain tumor image (ground truth) of a patient
    Ablation experimental results with different module settings
    Visualization results of two-dimensional segmentation of different models (green: edema; yellow: enhanced tumor; red: necrotic and non-enhancing tumor; green+yellow+red: whole tumor; yellow+red: tumor core)
    Visualization results of three-dimensional segmentation of different models (green: edema; yellow: enhanced tumor; red: necrotic and non-enhancing tumor; green+yellow+red: whole tumor; yellow+red: tumor core)
    Boxplot showing the segmentation results of different models on the three regions
    • Table 1. Experimental results of different designed modules

      View table

      Table 1. Experimental results of different designed modules

      ModuleMean Dice efficient /%Mean sensitivity /%Mean 95% Hausdorff distance /mm
      WTTCETWTTCETWTTCET
      UNet (base)89.7985.1380.9094.0885.4680.7713.617.475.45
      UNet+MR89.8986.4585.7192.1686.4587.137.265.485.14
      UNet+SC90.1187.3486.4993.9787.1688.935.477.084.93
      Proposed91.1387.4687.9895.4292.8787.484.435.342.53
    • Table 2. Experimental results comparison of different models

      View table

      Table 2. Experimental results comparison of different models

      ModelMean Dice coefficient /%Mean sensitivity /%Mean 95% Hausdorff distance /mmInference time /s
      WTTCETWTTCETWTTCET
      3DUNet89.7985.1380.9094.0887.4680.7713.617.475.453.56
      Attention UNet88.9882.3681.2391.4586.4576.5714.1210.178.535.13
      UNet++90.2784.0679.5594.2385.7780.4110.219.726.785.94
      ET-Net90.7886.3984.7392.5489.4686.156.575.493.194.42
      Point-UNet91.6187.0786.4291.6187.9487.189.167.138.757.43
      TransBTS91.1586.4386.4792.4194.9286.396.579.166.658.91
      Proposed91.1387.4687.9895.4292.8787.484.435.342.538.94
    Tools

    Get Citation

    Copy Citation Text

    Hengyi Tian, Yu Wang, Hongbing Xiao. Full‐Automatic Brain Tumor Segmentation Based on Multimodal Feature Recombination and Scale Cross Attention Mechanism[J]. Chinese Journal of Lasers, 2024, 51(21): 2107110

    Download Citation

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

    Category: Biomedical Optical Imaging

    Received: Apr. 16, 2024

    Accepted: Jul. 9, 2024

    Published Online: Oct. 31, 2024

    The Author Email: Wang Yu (wangyu@btbu.edu.cn), Xiao Hongbing (x.hb@163.com)

    DOI:10.3788/CJL240779

    CSTR:32183.14.CJL240779

    Topics