To build a compact model that meets the requirements of small medical devices, a lightweight self-distillation network (LSDNet) for cervical cell classification is developed in this paper. First, a multi-scale large kernel attention module is proposed to enhance extraction capability of multi-scale fine-grained features. Second, a deep layer selective aggregation structure is designed to selectively combine features at different levels in the network. Finally, the knowledge self-distillation is adopted to compress the network by inducing shallow classifier from different depths of the network, with the accuracy of the deepest classifier improved. The experimental results illustrate that the proposed LSDNet has the fewest parameters and achieves the most advanced classification performance compared to classical networks in recent publications.