<p>Accurate classification of cervical cell images is of great significance for early screening and diagnosis of cervical cancer. However, the diversity of cell morphology, the variation of staining methods, and the scarcity of high-quality labeled samples bring great challenges to traditional supervised learning methods. In this paper, we propose a structure-constrained network (SC-Net) for robust cervical cell representation learning. SC-Net imposes spatial structure consistency constraints on different enhanced views of the same cell through a structure-aware learning framework, and introduces Earth Mover’s Distance (EMD) as a metric for the alignment of structure distribution to promote the network to extract stable and discriminative features without supervision. We verified it on SIPaKMeD, a publicly available Pap smear dataset of isolated cervical cells, and obtained 97.73% classification accuracy. The excellent performance of SC-Net showed good robustness, which provided a new and effective solution for large-scale automatic analysis of clinical cervical cytology.</p>