The Combinations of Loss Functions and Schemes for Mammographic Classification

The analysis of mammographic images effectively is a burning question recently. So the research of binary classification whether a whole mammographic image has masses or not is proposed in this paper. The two main issues are the class imbalance of INbreast database and the effect of classification. To address these, focal loss, center loss and three kinds of schemes are proposed. Focal loss is put forward to address class imbalance by multiplying a modulating factor. Center loss concerns about enhancement of discriminative power by handling inter-class features, that is similar to the first two schemes. Moreover, the third scheme is used to solve the class imbalance with $L_{1}$ norm. The experimental analyses demonstrate the effectiveness of the combination of diverse loss functions and schemes. The convergence rate increases to varying degrees.