A Triplet Contrast Learning of Global and Local Representations for Unannotated Medical Images

Zhiwen Wei, Sungjoon Park, Jaeil Kim

Abstract: Recently, self-supervised learning(SSL) has shown its great potential in representation learning and been applied to various computer vision tasks. With the success of SSL, which showed performance improvement in natural images, SSL research is actively being conducted in medical image analysis. In this paper, we present a triplet network for the medical image representation learning to learn robust patterns of medical images against global and local changes by comparing latent feature distance between positive and negative pairs with anchors. This approach does not require large batches or the asymmetry of the network. It has been experimentally shown that the proposed method can outperform ImageNet pretrained models and the state-of-the-art SSL methods.