Search for a command to run...
Breast cancer (BC) is a deadly disease, killing millions of people every year. Developing automated malignant BC detection system applied on patient's imagery can help dealing with this problem more efficiently, making diagnosis more scalable and less prone to errors. Not less importantly, such kind of research can be extended to other types of cancer, making even more impact to help saving lives. Recent results on BC recognition show that Convolution Neural Networks (CNN) can achieve higher recognition rates than hand-crafted feature descriptors, but the price to pay is an increase in complexity to develop the system, requiring longer training time and specific expertise to fine-tune the architecture of the CNN. DeCAF (or deep) features consist of an in-between solution it is based on reusing a previously trained CNN only as feature vectors, which is then used as input for a classifier trained only for the new classification task. In the light of this, we present an evaluation of DeCaf features for BC recognition, in order to better understand how they compare to the other approaches. The experimental evaluation shows that these features can be a viable alternative to fast development of high-accuracy BC recognition systems, generally achieving better results than traditional hand-crafted textural descriptors and outperforming task-specific CNNs in some cases.