Multiview Deep Neural Networks for Tumor Classification in Breast Imaging

This study proposes a multiview convolutional neural network (CNN) for tumor classification in breast images. This CNN architecture receives multiple images simultaneously, specifically from three tumor regions of interest: internal, which contains only the tumor region; middle, which contains the transition zone between the tumor and the adjacent tissue; and external, which contains the peritumoral region. The multiview CNN was evaluated on mammography and breast ultrasound images and compared against a traditional single-view CNN. The multiview CNN achieves an AUC of 0.950 for mammograms, while the single-view CNN reaches 0.923. Likewise, the multiview CNN obtains an AUC of 0.936 for breast ultrasound, while its single-view counterpart attains 0.883. These results demonstrate the superiority of multiview CNNs in both imaging modalities because they obtain complementary information from several tumor regions of interest where radiologists usually focus their attention.