• Laser & Optoelectronics Progress
  • Vol. 59, Issue 2, 0200005 (2022)
Huan Zhang, Dawei Qiu, Yibo Feng, and Jing Liu*
Author Affiliations
  • College of Intelligence and Information Engineering, Shandong University of Traditional Chinese Medicine, Jinan , Shandong 250355, China
  • show less
    DOI: 10.3788/LOP202259.0200005 Cite this Article Set citation alerts
    Huan Zhang, Dawei Qiu, Yibo Feng, Jing Liu. Improved U-Net Models and Its Applications in Medical Image Segmentation: A Review[J]. Laser & Optoelectronics Progress, 2022, 59(2): 0200005 Copy Citation Text show less
    Schematic of U-Net architecture[6]
    Fig. 1. Schematic of U-Net architecture[6]
    Schematic representation of MDU-Net architecture[8]
    Fig. 2. Schematic representation of MDU-Net architecture[8]
    Schematic representation of R2U-Net architecture[13]
    Fig. 3. Schematic representation of R2U-Net architecture[13]
    Schematic representation of Bridged U-Net[14]
    Fig. 4. Schematic representation of Bridged U-Net[14]
    Schematic representation of M-Net architecture[15]
    Fig. 5. Schematic representation of M-Net architecture[15]
    Summary of U-Net model improvement
    Fig. 6. Summary of U-Net model improvement
    Normalization methods[38]
    Fig. 7. Normalization methods[38]
    Partial blood vessel region segmentation diagram[45]. (a) Original color fundus retinal images; (b) local fundus retinal images; (c) local standard retinal segmentation images; (d) local retinal segmentation result images
    Fig. 8. Partial blood vessel region segmentation diagram[45]. (a) Original color fundus retinal images; (b) local fundus retinal images; (c) local standard retinal segmentation images; (d) local retinal segmentation result images
    Three-dimensional segmentation results of different networks[51]
    Fig. 9. Three-dimensional segmentation results of different networks[51]
    Segmentation results of Base U-Net and BSU-Net [2]
    Fig. 10. Segmentation results of Base U-Net and BSU-Net [2]
    Segmentation results of brain tumor regions[58]
    Fig. 11. Segmentation results of brain tumor regions[58]
    PurposeImprovement measureAssociated network
    Avoiding overfitting

    1)Multiple types of dense connections

    2)Multiple technologies of data enhancement

    MDU-Net
    Reducing the number of parameters

    1)Inception module

    2)The global pooling layer

    3)Deep supervision

    4)Dense skip connection

    5)Full-scaled skip connection

    6)Skip connection using addition

    MultiResUNet;BSU-Net;GP-UNet;DENSE-Inception U-Net;UNet++;UNet3+;LadderNet;Bridged U-Net
    Focusing on effective features and suppressing irrelevant features

    1)Attention module

    2)SE module

    Attention U-Net;AnatomyNet;RA-UNet;ANU-Net
    Enhancing feature fusion

    1)Residual module

    2)Dense module

    3)Dense skip connection

    4)Full-scaled skip connection

    5)Feature pyramid

    6)Bidirectional feature network

    Vnet;MDU-Net;FD-UNet;Bridged U-Net;Dense Multi-path U-Net;UNet++;UNet3+;RA-UNet;DPSN;U-Det
    Speeding up convergence

    1)Residual module

    2)Dense skip connection

    3)Full-scaled skip connection

    4)Bridged U-Net

    5)Data normalization

    VNet;GP-UNet;Bridged U-Net;

    UNet++;nnUNet

    Enlarging receptive field

    1)Deformable convolution

    2)Dilated convolution

    DU-Net(Deformable U-Net);DMFNet;BSU-Net;3D-HDC-Unet
    Avoiding gradient vanishing or gradient explosion

    1)Residual module

    2)Attention residual module

    BSU-Net;RA-UNet;DENSE-Inception U-Net
    Table 1. Improvement measures of U-Net network for different purposes
    Area of segmentationMain image typeDifficultyImproved contentAssociated network
    Retinal vesselFundus color image

    1)The blood vessels are small

    2)Different shapes

    3)Accuracy of segmentation is low

    a)Adding deformable convolution or dilated convolution

    b)Adding up sampling feature channel

    c)Using attention mechanisms

    d)Adding inception module

    DU-Net;CASU;
    Pulmonary noduleCT image

    1)The aim area of segmentation is small

    2)The edges are blurry

    3)The contrast is low

    4)The grayscale is uneven

    5)Similar to tissues such as blood vessels in the essence of the lungs

    6)Shape heterogeneity is high

    a)Expanding to 3D

    b)Dense module

    c)Deep supervision

    d)Multi-scaled feature extraction

    e)Attention mechanisms

    3D ResUNet;

    CRF 3D U-Net;

    MSVNet;

    Double attention 3D U-Net;

    Liver tumorCT image

    1)Shape and size are irregular

    2)Similar to surrounding organs

    3)There are differences in grayscale values

    a)Using conditional random field(CRF)

    b)Adding dense module or inception module

    c)Using inverted residual bottleneck block(IRB block)

    3D UNet-C2;

    BSU-Net;

    LV-Net;

    Brain tumorMRI image

    1)Shape heterogeneity is high

    2)The structure of brain tissue is complex

    3)The boundaries are blurry

    4)Class imbalance is prominent

    a)Multiple technologies of data enhancement

    b)Multiple network synthesis

    c)Mixed dilated convolution

    d)Mixed loss function

    3D U-Net;

    3D-HDC-UNet;

    Table 2. Summary of image segmentation for various diseases
    Huan Zhang, Dawei Qiu, Yibo Feng, Jing Liu. Improved U-Net Models and Its Applications in Medical Image Segmentation: A Review[J]. Laser & Optoelectronics Progress, 2022, 59(2): 0200005
    Download Citation