• Optics and Precision Engineering
  • Vol. 31, Issue 6, 936 (2023)
Hong YU1,*, Renze LUO1, Chunmeng CHEN2, Xiang TANG3, and Renquan LUO1
Author Affiliations
  • 1College of Electrical Engineering and Information,Southwest Petroleum University, Chengdu60500, China
  • 2Department of Nuclear Medicine, The No. People’s Hospital of Yibin, Yibin644000, China
  • 3College of Computer Science, Southwest Petroleum University, Chengdu610500, China
  • show less
    DOI: 10.37188/OPE.20233106.0936 Cite this Article
    Hong YU, Renze LUO, Chunmeng CHEN, Xiang TANG, Renquan LUO. Bone scintigraphic classification method based on ACGAN and transfer learning[J]. Optics and Precision Engineering, 2023, 31(6): 936 Copy Citation Text show less
    ACGAN principle flow
    Fig. 1. ACGAN principle flow
    Overall flow chart of the method in this paper
    Fig. 2. Overall flow chart of the method in this paper
    Improved network models
    Fig. 3. Improved network models
    CA modules
    Fig. 4. CA modules
    Residual gated attention architecture
    Fig. 5. Residual gated attention architecture
    Dense residual attention module
    Fig. 6. Dense residual attention module
    CNN classification model
    Fig. 7. CNN classification model
    Comparison of MU-ACGAN training process
    Fig. 8. Comparison of MU-ACGAN training process
    Comparison of transfer learning
    Fig. 9. Comparison of transfer learning
    Bone images generated by MU-ACGAN
    Fig. 10. Bone images generated by MU-ACGAN
    Generative bone imaging and real bone imaging
    Fig. 11. Generative bone imaging and real bone imaging
    Similarity comparison
    Fig. 12. Similarity comparison
    Bone scintigraphy to identify heatmaps
    Fig. 13. Bone scintigraphy to identify heatmaps
    Confusion matrix
    Fig. 14. Confusion matrix
    IndexHealthMetastasesBenign changes
    Train/TestTrain/TestTrain/Test
    Quantity130/32101/2543/15
    Total16212658
    Ratio46.82%36.42%16.76%
    Table 1. Number and ratio of training and testing of bone imaging
    ClassHealthMetastasesBenign changes
    Quantity1 9501 515645
    Table 2. Amount of bone imaging data after the expansion
    Image typeHealthMetastasesBenign changes
    ACGAN0.458 50.516 00.506 8
    MU-ACGAN0.539 00.557 20.548 5
    Table 3. Quality assessment of bone imaging
    NumberAccuracyPrecisionRecallSpecificityF1-socre
    10.701 30.695 30.665 70.680 10.841 6
    20.727 30.730 10.687 80.708 30.845 2
    30.740 30.737 50.675 20.705 00.850 8
    40.779 20.803 20.692 80.743 90.868 8
    50.805 20.866 20.728 40.791 30.882 5
    Table 4. Classification effects of different data augmentation methods
    TypeAccuracyRecall
    Transfer learning0.805 20.728 4
    Dataset50.662 30.589 9
    Training directly0.597 40.527 6
    Table 5. Comparison of transfer learning strategies
    ModelsAccuracyRecallF1-socre
    VGG160.766 20.666 30.743 6
    DenseNet1210.805 20.755 00.767 7
    ResNext500.766 20.696 80.736 7
    Multi-scale CNN0.831 20.763 90.817 8
    Table 6. Experimental evaluation of multi-scale CNN model ablation
    ModelsAccuracyPrecisionRecallSpecificityF1-socre
    ResNet340.779 20.803 20.692 80.868 80.743 9
    MobileNetV30.779 20.785 70.745 90.875 60.765 2
    InceptionV30.792 20.847 70.750 60.874 20.796 2
    EfficientNet0.805 20.868 40.741 80.880 60.800 1
    Ours0.831 20.879 90.763 90.899 20.817 8
    Table 7. Comparison of the experimental results of different CNN models
    DatasetAccuracyRecallSpecificity
    Dataset40.831 20.763 90.899 2
    Dataset50.857 10.804 00.912 0
    Table 8. Evaluation metrics of different CNN classification models
    Hong YU, Renze LUO, Chunmeng CHEN, Xiang TANG, Renquan LUO. Bone scintigraphic classification method based on ACGAN and transfer learning[J]. Optics and Precision Engineering, 2023, 31(6): 936
    Download Citation