We present an application of deep Convolutional Neural Networks (CNN) for the detection and diagnosis of breast tumors. The images used in this study have been

extracted from the mini-MIAS database of mammograms. The proposed system has been implemented in three stages: (a) crop, rotation and resize of the original mammogram; (b) feature extraction using a pretrained CNN model (AlexNet and VGG); (c) training of a Support Vector Machine (SVM) at the classification task using the previously extracted features. In this research, the goal of the system is to distinguish between three classes of patients: those with benign, malign or without tumor. Experiments show that feature extraction using pretrained models provides satisfactory results, achieving a 64.52% test accuracy. This outcome could be improved via fine-tuning of the final layers or training the whole network parameters. The results of additional experiments using a sample of the Caltech-101 database, for which a 99.38% test accuracy was obtained, exhibit the relevance of the similarity between the data used to train the model and the particular application intended. Additionally, it is worth noting the impact of the data augmentation process and the balance of the number of examples per class on the performance of the system.