Local Cross-View Transformers and Global Representation Collaborating for Mammogram Classification

When analyzing screening mammography images, radiologists compare multiple views of the same breast to help improve the detection rate of lesions and reduce the incidence of false-positive results. Therefore, to make the deep learning-based mammography computer-aided detection/diagnosis (CAD) system meet the radiologists’ requirements for accuracy and generality, the construction of deep learning models needs to mimic manual analysis and consider the correlation between different views of the same breast. In this paper, we propose the Local Cross-View Transformers and Global Representation Collaborating for Mammogram Classification (LCVT-GR) model. The model uses different view images to train in an end-to-end manner. In this model, the global and local representations of mammogram images are analyzed in parallel using the global-local parallel analysis method. To validate the effectiveness of our method, we conducted comparison experiments and ablation experiments on two publicly available datasets, Mini-DDSM and CMMD. The results of the comparison experiments show that our method achieves better results compared with existing advanced methods, with greater improvements in both AUC-ROC and AUC-PR assessment metrics. The results of the ablation experiments show that our model architecture is scientific and effective and achieves a good trade-off between computational cost and model performance.