• Optics and Precision Engineering
  • Vol. 31, Issue 18, 2765 (2023)
Ping XIA1,2, Guangyi ZHANG1,2, Bangjun LEI1,2,*, Yaobing ZOU1,2, and Tinglong TANG1,2
Author Affiliations
  • 1Hubei Key Laboratory of Intelligent Vision based Monitoring for Hydroelectric Engineering, Three Gorges University, Yichang443002, China
  • 2College of Computer and Information Technology, Three Gorges University, Yichang44300, China
  • show less
    DOI: 10.37188/OPE.20233118.2765 Cite this Article
    Ping XIA, Guangyi ZHANG, Bangjun LEI, Yaobing ZOU, Tinglong TANG. Polyp image segmentation based on multi-scale ResNeSt-50 aggregation network and message passing[J]. Optics and Precision Engineering, 2023, 31(18): 2765 Copy Citation Text show less
    Model of this paper
    Fig. 1. Model of this paper
    Split-attention in ResNeSt Block
    Fig. 2. Split-attention in ResNeSt Block
    ResNeSt-50 encoding structure of this paper
    Fig. 3. ResNeSt-50 encoding structure of this paper
    RFB and dense aggregation module
    Fig. 4. RFB and dense aggregation module
    Message propagation of confidence level
    Fig. 5. Message propagation of confidence level
    Segmentation result of testing on the ETIS-LaribPolypDB dataset
    Fig. 6. Segmentation result of testing on the ETIS-LaribPolypDB dataset
    Segmentation result of testing on the Kvasir-SEG[34] dataset
    Fig. 7. Segmentation result of testing on the Kvasir-SEG34dataset
    Segmentation result of testing small lesion on ETIS-LaribPolypDB[1]dataset
    Fig. 8. Segmentation result of testing small lesion on ETIS-LaribPolypDB1dataset
    Comparison of the effect of the model used in this paper and PraNet model in segmenting minimal lesions
    Fig. 9. Comparison of the effect of the model used in this paper and PraNet model in segmenting minimal lesions
    Segmentation result of testing on the ETIS-LaribPolypDB[1]dataset
    Fig. 10. Segmentation result of testing on the ETIS-LaribPolypDB1dataset
    数据集Kvasir-SEG34CVC-ClinicDB33ColonDB35ETIS-LaribPolypDB 1CVC-30036
    数 量1 00061238019660
    图像大小500+×500+384×288574×5001 225×966574×500
    Table 1. Five datasets used in this paper
    训练集测试集
    实验AKvasir-SEG(900),CVC-ClinicDB(550)Kvasir-SEG(100)
    实验BKvasir-SEG(900),CVC-ClinicDB(550)CVC-ClinicDB(62)
    实验CKvasir-SEG(900),CVC-ClinicDB(550)ColonDB(380)
    实验DKvasir-SEG(900),CVC-ClinicDB(550)ETIS-LaribPolypDB (196)
    实验EKvasir-SEG(900),CVC-ClinicDB(550)CVC-300(60)
    Table 2. Allocation of training dataset and test dataset of experiment
    模 型测试集mDicemIoUSmeasureMAE
    模型aColonDB0.5120.4440.7120.061
    模型b0.6080.5230.7600.052
    模型c0.6760.6060.7990.043
    模型d0.7570.675n/an/a
    模型a

    ETIS-

    LaribPolypDB

    0.3980.3350.6840.036
    模型b0.5220.4470.7380.026
    模型c0.6140.5470.7810.028
    模型d0.7060.628n/an/a
    Table 3. Ablation experiment of ResNeSt-50 module(training on Kvasir-SEG and ClinicDB dataset)
    mDicemIoUSmeasureMAE
    未采用多尺度训练0.8880.8270.9080.031
    使用多尺度训练0.9130.8580.9250.023
    Table 4. Effect of multi-scale training on model segmentation
    EpochmDicemIoUSmeasureMAE
    Adam200.9040.8510.9200.027
    RAdam300.9130.8580.9250.023
    Table 5. Effect of RAdam optimizer on model segmentation
    mDicemIoUSmeasureMAE
    本文编码-解码器0.9130.8580.9250.023
    本文编码-解码器+TTA0.9150.8610.9250.023
    本文模型0.9160.8630.9210.023
    Table 6. Trained on CVC-ClinicDB33,Kvasir-SEG34dataset, compare the test results on Kvasir-SEG34dataset
    ColonDB35ETIS-LaribPolypDB1CVC-30036
    mDicemIoUmDicemIoUmDicemIoU
    本文编码-解码器0.7570.6750.7060.6280.8580.781
    本文编码-解码器 +TTA0.7830.7000.7640.6860.8690.797
    本文模型0.7860.7030.7700.6910.8710.798
    Table 7. Trained on the CVC-ClinicDB33,Kvasir-SEG34dataset, respectively compare the test results on ColonDB35,ETIS-LaribPolypDB1,CVC-30036
    mDicemIoUSmeasureMAE
    U-Net90.8180.7460.8580.055
    U-Net++100.8210.7430.8620.048
    ResUnet++110.7340.639n/an/a
    SFA380.7230.6110.7820.075
    PraNet120.8980.8400.9150.030
    本文方法0.9160.8630.9210.023
    Table 8. Trained on CVC-ClinicDB33,Kvasir-SEG34dataset, compare the test results on Kvasir-SEG34dataset
    mDicemIoUSmeasureMAE
    U-Net90.0550.0460.5170.021
    U-Net++100.0930.0780.5450.018
    SFA380.0120.0060.4250.130
    ResUnet++110.0160.0090.5000.011
    PraNet120.0990.0730.5150.074
    本文方法0.4870.4010.7340.003
    Table 9. Trained on CVC-ClinicDB,Kvasir-SEGdataset, compare the test results on ETIS-LaribPolypDBdataset
    ColonDB35ETIS-LaribPolypDB1CVC-30036
    mDiceMIoUmDiceMIoUmDiceMIoU
    U-Net90.5120.4440.3980.3350.710.627
    U-Net++100.4830.4100.4010.3440.7070.624
    ResUnet++11n/an/an/an/an/an/a
    SFA380.4690.3470.2970.2170.4670.329
    PraNet120.7090.6400.6280.5670.8710.797
    本文方法0.7860.7030.7700.6910.8710.798
    Table 10. Trained on the CVC-ClinicDB33,Kvasir-SEG34dataset, respectively compare the test results on ColonDB35,ETIS-LaribPolypDB1,CVC-30036dataset
    Ping XIA, Guangyi ZHANG, Bangjun LEI, Yaobing ZOU, Tinglong TANG. Polyp image segmentation based on multi-scale ResNeSt-50 aggregation network and message passing[J]. Optics and Precision Engineering, 2023, 31(18): 2765
    Download Citation