Attention-based skin lesion recognition

Skin cancer is one of the most common types of cancers in the world and is a big concern for people’s health. In recent years, automatic algorithms to recognize skin cancers from dermoscopy images have gained lots of popularity, especially deep-learning-based methods. In this thesis, we propose an attention-based deep learning model for skin cancer recogni- tion. The attention modules, which are learned together with other network parameters, estimate attention maps that highlight image regions of interest that are relevant to lesion classification. These attention maps provide a more interpretable output as opposed to only outputting a class label. Additionally, we propose to utilize prior information by regulariz- ing attention maps with regions of interest (ROIs) (e.g., lesion segmentation or dermoscopic features). To our knowledge, we are the first to introduce an end-to-end trainable attention module with regularization for skin cancer recognition. We provide both quantitative and qualitative results on public datasets to demonstrate the effectiveness of our method. Experiments show that: (1) the attention module is capable of ruling out irrelevant areas in the image; (2) when the proposed attention regularization terms are applied, both the classification performance and the attention maps can be further refined; (3) the attention regularization is quite robust and flexible in that it can take advantage of sparse or even imperfect ROI maps. The code of this work is released at https://github.com/SaoYan/IPMI2019-AttnMel.