Breast Cancer Subtype Classification Based on PET/CT Bimodal Imaging Feature Fusion

In this paper, the powerful feature extraction capability of convolutional neural networks is used to extract advanced latent features of breast lesions, and the classification of breast cancer subtype is realized based on the features extracted from PET/CT bimodal images. The DenseNet network model is used to construct the PET/CT imaging feature extraction network. Since the PET/CT bimodal imaging data from breast cancer patients is very limited, the transfer learning method is adopted to solve the problem of insufficient data and prevent overfitting during training. In order to combine PET/CT bimodal imaging features, two feature fusion approaches are proposed, which are offline feature fusion and online feature fusion. For offline feature fusion, principal component analysis is employed to reduce the dimensionality of the deep features extracted by the feature extraction network from individual modalities. The reduced features are then concatenated and forwarded into a classifier for breast cancer subtype classification. For the online feature fusion, the feature maps output by the PET/CT feature extraction networks are directly concatenated and then input into the classification subnetwork for breast cancer subtype classification. The validity is evaluated using clinical data in practice. The PET/CT dual-modality features can improve the accuracy of breast cancer classification, and the experimental results show that the accuracy of online feature fusion is the highest.