Generative Models for Class Imbalance Problem on BreakHis Dataset: A Case Study

In real-world classification tasks, it is common to find class imbalance issues in the training datasets, i.e. an unequal number of examples among the different classes. The class imbalance problem biases the performance of predictive models by overlooking minority classes; this is because predictive models employ learning rules with accuracy-based cost functions, thus favoring majority classes. In this work, the class imbalance issue is tackled through generative models, using the BreakHis dataset, a histopathologic image set intended for breast cancer classification, as a case study. The BreasHis’ minority class is balanced by adding synthetic images obtained by means of different generative methods, including variational autoencoders and two different generative adversarial networks. The quality of the image sets created by the different generative models, and their effects in balancing the BreakHis dataset, are evaluated through several quantitative metrics computed from classification tasks. Statistical analysis is performed and the results indicate that the DCGAN network is superior to the other evaluated models.