• Acta Optica Sinica
  • Vol. 40, Issue 18, 1810003 (2020)
Hong Huang1、*, Chao Peng1, Ruoyu Wu1, Junli Tao2, and Jiuquan Zhang2
Author Affiliations
  • 1Key Laboratory of Optoelectronic Technology & Systems, Ministry of Education, Chongqing University, Chongqing 400044, China;
  • 2Department of Radiology, Chongqing University Cancer Hospital, Chongqing 400030, China
  • show less
    DOI: 10.3788/AOS202040.1810003 Cite this Article Set citation alerts
    Hong Huang, Chao Peng, Ruoyu Wu, Junli Tao, Jiuquan Zhang. Self-Supervised Transfer Learning of Pulmonary Nodule Classification Based on Partially Annotated CT Images[J]. Acta Optica Sinica, 2020, 40(18): 1810003 Copy Citation Text show less
    Schematic of self-supervised learning principle
    Fig. 1. Schematic of self-supervised learning principle
    Transformation method of lung CT images. (a) Original image; (b) nonlinear transformation; (c) local pixel change; (d) external pixel change; (e) internal pixel change
    Fig. 2. Transformation method of lung CT images. (a) Original image; (b) nonlinear transformation; (c) local pixel change; (d) external pixel change; (e) internal pixel change
    Flow chart of proposed algorithm
    Fig. 3. Flow chart of proposed algorithm
    Loss-accuracy graph of proposed model when FC is 2
    Fig. 4. Loss-accuracy graph of proposed model when FC is 2
    ROC curves of different classification models when FC is 2
    Fig. 5. ROC curves of different classification models when FC is 2
    NumberModelACCSenSpeAUC
    Scratch model0.815±0.0180.698±0.0650.908±0.0460.879±0.022
    FC is 1Self-supervised model0.828±0.0010.684±0.1130.903±0.0430.895±0.008
    Proposed model0.843±0.0390.754±0.1030.914±0.0200.924±0.015
    Scratch model0.817±0.0180.751±0.0800.867±0.0540.870±0.026
    FC is 2Self-supervised model0.864±0.0130.812±0.0480.903±0.0500.915±0.009
    Proposed model0.886±0.0090.839±0.0540.920±0.0440.929±0.016
    Scratch model0.819±0.0480.767±0.0110.861±0.0650.878±0.039
    FC is 3Self-supervised model0.830±0.0250.768±0.0680.877±0.0570.905±0.030
    Proposed model0.841±0.0490.855±0.0300.831±0.1000.906±0.038
    Table 1. Comparison of classification performance under different number of fully connected layers
    NumberScratch modelSelf-supervised modelProposed model
    FC is 15.044.314.34
    FC is 25.084.694.68
    FC is 35.344.875.39
    Table 2. Comparison of running time of different models under different number of fully connected layerss
    AlgorithmACCSenSpeAUC
    SVM0.746±0.0320.660±0.0610.814±0.0270.802±0.025
    3D CNN0.811±0.0540.740±0.1160.868±0.0400.872±0.055
    3D ResNet0.826±0.0400.758±0.1210.880±0.0500.906±0.026
    3D DenseNet0.851±0.0280.847±0.0340.854±0.0360.908±0.010
    Proposed algorithm0.886±0.0090.839±0.0540.920±0.0440.929±0.016
    Table 3. Comparison of classification performance of different algorithms
    AlgorithmSVM3D CNN3D ResNet3D DenseNetProposed algorithm
    Time /s0.085.307.1915.234.68
    Table 4. Comparison of running time of different algorithms
    Hong Huang, Chao Peng, Ruoyu Wu, Junli Tao, Jiuquan Zhang. Self-Supervised Transfer Learning of Pulmonary Nodule Classification Based on Partially Annotated CT Images[J]. Acta Optica Sinica, 2020, 40(18): 1810003
    Download Citation