Deep Metric Learning by Exploring Confusing Triplet Embeddings for COVID-19 Medical Images Diagnosis

Because the COVID-19 virus is highly transmissible, leading to a worldwide increment of new infections and deaths daily, the development of an automated tool to identify COVID-19 using CT images has attracted much attention. Significantly, deep metric learning can be deployed to cluster and classify the fine-grained CT images, which aims to learn a mapping from the original objects to a discriminative feature embedding space. Previous deep metric learning works have been proposed to construct various structures of loss, mine hard samples, or introduce regularization constraints, \etc. In general, traditional loss functions of deep metric learning methods are based on constraining the distance of the triplet embeddings in the feature space. Instead of focusing on the previous research directions, in this work, we pay attention to exploring confusing triplet embeddings, for the reason that confusing triplet embeddings perform a side effect on the majority of deep triplet-based metric learning methods. By considering the spatial relation of triplet embedding, and conducting theoretical analysis in the feature space, we propose an approach to recognize the confusing triplet embeddings and construct a Confusing Triplet Embedding Learning (CTEL) method by adding a hard constraint on the confusing triplet embeddings. The extensive experiments indicate that our proposed CTEL method achieves more excellent performance on two medical CT image datasets and two fine-grained standard image datasets compared with many state-of-the-art methods.