This paper presents an in‐depth study on the contribution and integration of attention mechanisms into deep learning architectures for melanoma classification. Indeed, the concept of attention helps guide the learning process to focus on some parts of the input image deemed to be the most significant. Nevertheless, to the best of our knowledge, a study on how and where to integrate such mechanisms has never been conducted in the context of melanoma classification. Consequently, we propose such a study in three main stages. First, we propose an improvement in the transfer learning process for the specific case of melanoma classification. Then, we develop a basic variant of our architecture allowing to integrate an additive attention mechanism in an adapted ResNet network. We finally investigate the impact of the number of mechanisms and their location in the melanoma classification architecture. Better results are reported on two standard datasets.

No abstract