Interactive Attention Sampling Network for Clinical Skin Disease Image Classification

Skin disease is one of the global burdens of disease, and affects around 30% to 70% individuals worldwide. Effective automatic diagnosis is indispensable for doctors and patients. Compared with dermoscopic imaging, using clinical images captured by a portable electronic device (e.g. a mobile phone) is more available and low-cost. However, the existing large clinical skin-disease image datasets do not have the spatial annotation information, thus posing challenges for localizing the skin-disease regions and learning detailed features. To address the problem, we propose the Interactive Attention Sampling Network (IASN) which automatically localizes the target skin-disease regions and highlight the regions into high resolution. Specifically, the top-K local peaks of the class activation maps are collected, which indicate the key clues of skin-disease images. Then the features of the local peaks are interacted with the features of the surrounding context. With the guidance of the interactive attention maps, the non-uniform sampled images are generated, which facilitate the model to learn more discriminative features. Experimental results demonstrate that the proposed IASN outperforms the state-of-the-art methods on the SD-198 benchmark dataset.