MLSE-FML: Mammogram classification with multi-level semantic extraction and focal margin loss

Mammography screening is a powerful tool for detecting early breast cancer. Although deep learning (DL)-based mammography screening models have demonstrated promising results for diagnosing breast cancer, the development of them faces two challenges. One challenge is the information loss of local lesion details caused by the pooling and convolutional layers in convolutional neural networks. The other challenge is the ambiguous inter-class difference and large intra-class variance due to the heterogeneity of breast lesions. In the present study, a model called multi-level semantic extraction and focal margin loss (MLSE-FML) is proposed to address the two challenges, and it contains a multi-level semantic extraction (MLSE) module and a focal margin loss (FML) function. The MLSE module is used to fuse semantically strong multi-level features for alleviating the influence of information loss of breast lesions on the models. The FML function is developed as a margin-based loss function with a proposed adaptive probability-aware margin to impose different constraints on training samples for extracting discriminative features. Experimental results demonstrate that the performance of the proposed model (AUC = 0.777 on the VinDr-Mammo dataset, AUC = 0.826 on the CBIS-DDSM dataset, and AUC = 0.969 on the INbreast dataset) is better than that of several state-of-the-art models on the three public datasets. Additionally, the heat-maps demonstrate that the model has the ability to localize the lesion areas. Therefore, the model has the potentials to be used in clinical practice for diagnosing breast cancer. The code of the proposed model is available at https://github.com/85675/MLSE-FML .