DeepMammo: Deep Transfer Learning for Lesion Classification of Mammographic Images

Recently, impressive results have been provided by pre-trained convolutional neural networks combined with the transfer learning technique. They have quickly become a great option to classify general image datasets. However, to the best of our knowledge, the majority of works do not explore if these pre-trained architectures are well-suited to specific contexts like the medical image domain (e.g. breast lesions). We focus on breast lesions, because it is one of the most common types of cancer affecting women worldwide, and its early diagnosis is crucial to the success of the treatment. In this paper, we propose a methodology capable of analyzing different approaches, regarding description (e.g through handcrafted and deep features) and classification (e.g. through end-to-end networks and traditional classifiers) to automatically answer the best tuning according to a given mammographic image dataset. Our methodology can also apply a data augmentation method to improve the learning of the networks. It increases the number of training samples to cope with unbalancing and overfitting problems that are intrinsic to the breast lesion classification task. We validate our methodology on public image datasets and our results show classification accuracies of up to 94.34%.