Considering breast density for the classification of benign and malignant mammograms

Background and objective: Mammography plays a crucial role in breast cancer screening because it can be used to diagnose a breast mass and breast calcification region early. Mammogram has some drawbacks; for example, detecting breast cancer masses in extremely dense breast tissue is considerably difficult. Thus, this study combined breast density with benign and malignant masses to classify breast tumors. Method: This study used data augmentation to preprocess mammograms and then convolutional neural network (CNN) models namely AlexNet, DenseNet, and ShuffleNet to classify the images. The accuracies of the three models before and after data augmentation were compared. Results: Before data augmentation, the accuracies for training and testing sets of AlexNet were 40.47 % and 38.57 %, respectively, those of DenseNet were 90.00 % and 46.39 %, respectively, and those of ShuffleNet were 96.48 % and 42.86 %, respectively. After data augmentation, the accuracies for training and testing sets of AlexNet were 99.35 % and 95.46 %, respectively, those of DenseNet were 99.91 % and 99.72 %, respectively, and those of ShuffleNet were 99.85 % and 97.84 % respectively. Conclusions: Data augmentation allows the model to learn more pictures of different situations and angles to accurately classify new images. Although the accuracy of DenseNet was the highest, the execution time required was considerably longer than that required for ShuffleNet. Considering the accuracy and execution time, ShuffleNet is recommended to be used for the classification of benign and malignant mammography breast tumors.