Investigating the Effect of Various Augmentations on the Input Data Fed to a Convolutional Neural Network for the Task of Mammographic Mass Classification

Along with the recent improvement in medical image analysis, exploring deep learning based approaches in the context of mammography image processing has become more realistic. In this paper, we concatenate on both conventional machine learning and deep learning approaches to classify mass abnormalities in mammographic images. Using a deep convolutional neural network (CNN) architecture, the effect of performing various augmentation approaches on the raw pre-detected masses fed to the network is investigated. We propose an extended augmentation method, specific filter bank responses and also a texton-based approach to generate characteristic filtered features for various types of mass textures and eventually use the resulting image data as input for training the CNN. Evaluating our proposed techniques on the DDSM dataset, we show that mammographic mass classification can be tackled effectively by employing an extended augmentation scheme. We obtained 87% accuracy which is comparable to the currently reported results for this task.