Classification of Histopathological Breast Cancer Images Using Convolutional Neural Networks

Abstract

Breast cancer is the most common cancer that affects women both in the United States and worldwide. The time in which breast cancer is diagnosed is crucial, as the survival rate is much higher when the cancer is caught in its early stages. The BreakHis Dataset from the University of Parana in Brazil provides over 7000 microscopic images of benign and malignant breast tumors upon which classification tasks can be performed. (Spanhol et al., 2016) The tumors are also labeled with their specific type, something that is unique to this dataset, resulting in 8 subcategories. This paper attempts different classification tasks on this dataset by building convolutional neural network (CNN) models, a powerful image classification architecture. We find that for the two class model, we can approach average accuracy rates near 87% with relatively simple CNN models. However, we find that it is much more difficult to successfully classify the subcategory of tumor, reaching only about 54% accuracy rates. This leaves room for further study using more powerful deep neural networks to achieve high accuracy.