• Laser & Optoelectronics Progress
  • Vol. 58, Issue 4, 0410007 (2021)
Wenhui Xu, Yijian Pei*, Donglin Gao, Jiude Zhu, and Yunkai Liu
Author Affiliations
  • School of Information, Yunnan University, Kunming, Yunnan 650500, China
  • show less
    DOI: 10.3788/LOP202158.0410007 Cite this Article Set citation alerts
    Wenhui Xu, Yijian Pei, Donglin Gao, Jiude Zhu, Yunkai Liu. Mass Classification of Breast Mammogram Based on Attention Mechanism and Transfer Learning[J]. Laser & Optoelectronics Progress, 2021, 58(4): 0410007 Copy Citation Text show less
    Workflow of the proposed method
    Fig. 1. Workflow of the proposed method
    Mammograms and the pixel-level lesion labels of different views (L-CC, L-MLO, R-CC and R-MLO) from CBIS-DDSM dataset
    Fig. 2. Mammograms and the pixel-level lesion labels of different views (L-CC, L-MLO, R-CC and R-MLO) from CBIS-DDSM dataset
    Images from different datasets. (a)Local breast mass patch image from dataset P; (b)whole breast mammogram from CBIS-DDSM dataset;(c)global breast mammogram from dataset G
    Fig. 3. Images from different datasets. (a)Local breast mass patch image from dataset P; (b)whole breast mammogram from CBIS-DDSM dataset;(c)global breast mammogram from dataset G
    Residual block structure in ResNet50 Network
    Fig. 4. Residual block structure in ResNet50 Network
    Structure of channel attention module
    Fig. 5. Structure of channel attention module
    Structure of spatial attention module
    Fig. 6. Structure of spatial attention module
    Mammogram from dataset G and heat maps of the corresponding Grad-CAM. (a)Sample 1(benign); (b)sample 2(benign);(c)sample 3(malignant);(d)sample 4(malignant);(e)sample 5(malignant)
    Fig. 7. Mammogram from dataset G and heat maps of the corresponding Grad-CAM. (a)Sample 1(benign); (b)sample 2(benign);(c)sample 3(malignant);(d)sample 4(malignant);(e)sample 5(malignant)
    ModelSensitivitySpecificityAccuracyAUC
    VGG16VGG16+CBAM0.76190.76490.77480.80730.76980.78310.83950.8503
    Inception V3Inception V3+CBAM0.76870.76990.77920.78060.77510.77720.84030.8477
    ResNet500.75510.78350.77240.8348
    Proposed network0.79930.80740.80420.8607
    Table 1. Comparison of classification performance of different network models on patch dataset P
    MethodTransfer learningSensitivitySpecificityAccuracyAUC
    VGG16ImageNetPatch data0.68350.71430.72720.73800.70970.72880.75990.7749
    VGG16+CBAMImageNetPatch data0.71940.72780.71770.74080.71840.73350.78850.7913
    Inception V3ImageNetPatch data0.66190.71220.67940.74160.67240.72980.75410.7941
    Inception V3+CBAMImageNetPatch data0.68350.74820.73210.69860.71260.71840.79540.7656
    ResNet50ImageNetPatch data0.71220.74820.74640.70370.73270.73410.78810.7929
    Proposed networkImageNetPatch data0.73500.76980.74560.74160.74140.75290.79820.8081
    Table 2. Comparison of classification performance of different network models and different transfer learning methods on global breast dataset G
    Wenhui Xu, Yijian Pei, Donglin Gao, Jiude Zhu, Yunkai Liu. Mass Classification of Breast Mammogram Based on Attention Mechanism and Transfer Learning[J]. Laser & Optoelectronics Progress, 2021, 58(4): 0410007
    Download Citation