Table 1.
Data distribution.
Fig 1.
The overall structure of the MVNN model.
Two multi-scale attention DenseNets are used to extract the features of mammograms from two views, then a fully connected layer is used to fuse these features, finally realize the classification of mammograms in two stages.
Fig 2.
Overall structure of CSAM with parallel channel attention and spatial attention.
Fig 3.
The basic structure of Dense Block.
B: Building blocks of Dense Blocks for extracting features of feature map Xi. T: Transaction layer, which is used between two dense blocks to reduce the dimension of output features.
Fig 4.
Structure of multi-scale convolution module with CSAM attention module added.
Table 2.
Parameters of the multi-scale attention DenseNet with a depth of 186.
Fig 5.
The performance comparison between two single-view networks and the proposed MVNN in the two-stage classification tasks.
(a) Accuracy, (b) Sensitivity.
Fig 6.
Two popular multi-scale convolution modules.
(a) An ordinary multi-scale convolution module, (b) multi-scale convolution module inspired by Inception structure.
Fig 7.
ROC curves of different multi-scale convolution modules and original DenseNet.
(a) ROC curve of normal and abnormal classification, (b)ROC curve of Benign and Malignancy classification.
Table 3.
Parameters of the multi-scale attention DenseNet with a depth of 186.
Fig 8.
Visualization results of the model with different convolution module in normal and abnormal classification task.
This case is malignant, and P denotes the corresponding softmax score in the benign and malignant classification task.
Fig 9.
ROC curves of different attention modules in two-stage classification tasks.
(a) Normal and Abnormal classification, (b) Benign and Malignancy Classification.
Table 4.
Influence of different attention modules on model classification performance.
Fig 10.
Visualization results of the model with different attention module in normal and abnormal classification task.
This case is benign, and P denotes the corresponding softmax score in the benign and malignant classification task.
Table 5.
Diagnostic accuracy of mass lesions and calcified lesions in two classification tasks.
Table 6.
Test performance of the MVNN model.
Table 7.
Compare with State-of-the-art.