DOCToR: The Role of Deep Features in Content-Based Mammographic Image Retrieval
Rafael S. Bressan, Daniel H. A. Alves, Lucas M. Valerio, Pedro H. Bugatti, Priscila T. M. Saito · 2018
Nowadays, deep features, obtained from a variety of deep learning architectures, play an important role in several real problems. It is know that transfer learning strategies could be employed to take advantage of such deep features trained under a general context (e.g. ImageNet). However, to the best of our knowledge, the majority of works focus on similar contexts to accomplish such transfer strategies. Thus, in this work we analyze the role of deep features in content-based medical image retrieval, and demonstrate that it is possible to make use of transfer learning from a general context to a specific medical context, like the content-based mammographic image retrieval. To do so, we evaluated several hand-crafted features against deep features acquired from state-of-the-art deep architectures through transfer learning. Extensive experiments on challenging public mammographic image datasets testify that the generalized deep features are able to improve in a great extend the precision of similarity queries both in the traditional process and applying query refinement strategies.