Mass is one important suspicious object for breast cancer diagnosis in mammograms. Computer-aided detection (CAD) based on fully supervised deep learning achieves high performance for mass detection in mammograms. The lack of fine-grained expert labels becomes the bottleneck for the large-scale application of CAD to achieve detection in mammograms. Weakly supervised methods provide a solution to tackle the annotation problems, including in the application for mass detection. However, previous works face the problem of insufficient localization information, which affect the ability of mass detection. In this paper, we propose a multi-view enhancing mass detection network (MVMDNet) with dual view inputs that contains craniocaudal (CC) and mediolateral oblique (MLO) view of mammograms, where different view features are interacted and fused to enhance localization information. After backbone extracts dual view features, the multi-view enhancement is contributed by following modules: First, by Special Correlation Attention (SCA) module, the correlation between dual views features, which contains extra localization information, is extracted to generate the auxiliary features. Second, by Sigmoid Weighted Fusion module, examined features are generated from the fusion of auxiliary feature and the main diagnostic CC view feature, where the weights of fusion are automatically learned in training. Examined features are conducted a binary classification training by image-level labels that indicate whether masses are contained in mammograms. Aiming to realize the detection, Class Activating Mapping-based detection module catch the network’s attention regions on examined features to generate the detection boxes for breast mass. We conduct experiments on both in-house dataset and public INbreast dataset. The recall rate of 0.96±0.02 on public INbreast dataset and the recall rate of 0.92 on in-house dataset demonstrate the performance of MVMDNet is similar to the fully-supervised methods and achieves the state-of-the-art among weakly-supervised methods. This performance on the public INbreast dataset also indicates a robust generalization ability.