A Modified Relation and Margin-Based Deep Learning Network for Automatic Breast Cancer Detection

Breast cancer is the second leading cause of cancer deaths among women and one of the most frequently diagnosed cancers. Hence, early and accurate detection of breast cancer is quite challenging. Mammography and ultrasonography are two commonly used screening technologies for detecting breast cancer to diminish the mortality rate. Recently, several researchers have come up with computer-aided designs (CADs) to aid radiologists and improve detection efficacy. In this work, a modified Relation and Margin Network (MReMarNet) is presented for efficient breast cancer detection. The proposed model tries to improve the intraclass compactness and the inter-class separability for the classification of small sample datasets. Moreover, a relation unit (RU) and a fully connected (FC) unit with the cross-entropy loss are employed simultaneously for feature learning and decision boundary-based classification, respectively. The coupled benefits of intra-class compactness provided by RU and inter-class separability provided by the FC branch make the system more efficient. Two publicly available datasets, mini-DDSM (mammogram) and BUSI (ultrasound), are used for the experiment. The proposed breast cancer identification approach outperforms other networks with the accuracy of 98.75% and 95.77% and 98.00% for mini-DDSM, BUSI and BUS2 datasets, respectively, to detect malignancy.