Abstract: Objective The MMV-Net,a deep learning framework based on mammogram multiple views,was constructed to evaluate the classification performance of the model for benign and malignant masses. Methods A retrospective analysis was conducted on a dataset of 1 585 breast X-ray images from Harbin Medical University Cancer Hospital from 2018 to 2020,including 806 benign cases and 779 malignant cases.The dataset was divided into the training set(n=1 268)and the test set(n=317)according to an 8∶2 ratios,and the training set was stratified according to the 5-fold cross validation.The integrated DDSM dataset and INBreast dataset were used as external test sets(n=1 645)to evaluate the model performance.Each case in the input layer contained 4 views.The MMV-Net model was constructed by removing the last two layers of the ResNet22 network structure and adding an average pooling layer as the feature extraction layer,as well as fully connection layer and softmax activation function as the decision layers.Bayesian hyperparameter optimization was used.The performance of MMV-Net,MFA Net,and ensemble inception V4 models in AUC values,accuracy,precision,recall and F1 scores were compared. Results The AUC values of MMV-Net model for distinguishing benign and malignant masses on the test set were 0.913,0.882 for MFA-Net,and 0.865 for inception V4.The accuracy and precision evaluation metrics of the MMV-Net model were also higher than the other two models. Conclusion The deep learning MMV-Net model based on multiple views of mammogram is helpful for the classification of benign and malignant breast masses.

Key words: Deep learning, MMV-Net model, Mammography, Tumor, Classification