Deep Semantics-Preserving Hashing Based Skin Lesion Image Retrieval

This study proposes a content-based pigmented skin lesion image retrieval scheme on semantic hash clustering on the output of the deep neural networks. The skin lesion images are acquired with standard digital cameras or mobile phones. To retrieval skin lesion images efficiently online, semi-supervised deep convolutional neural network incorporated with hash functions jointly learn feature representations, for preserving similar semantics between skin lesion images, and mappings to hash codes. The target candidates are clustered by Affinity Propagation (AP) for ranking, which are selected among the outputs of layer F7 based on the Hamming distance of their semantic hash codes. Experiments on 4 disease categories of pigmented skin lesions of a set of 239 images yielded a specificity of 93.4% and a sensitivity of 80.89%.