Understanding the robustness of deep neural network classifiers for breast cancer screening