A lens-free on-chip digital holographic microscope (LFOCDHM) is essential for a variety of biomedical applications such as cell cycle assays, drug development, digital pathology, and high-throughput biological screening. However, due to the unit magnification configuration of the lens-free system, the field-of-view (FOV) contains over a hundred times more cells than a conventional 10× microscope objective. Consequently, the segmentation process becomes labor-intensive and time-consuming due to the complex and variable morphology of cells within the large FOV. To address this issue, numerous deep learning-based cell segmentation methods have been proposed. Nevertheless, convolutional neural networks, limited by their localized receptive field, are unsuitable for segmenting and processing large FOV imaging results from LFOCDHM. Therefore, we propose a high-throughput live cell analysis processing method called Swin Transformer U-Net (STU-Net). Based on the reconstructed phase results, a shift window is utilized to compute the self-attention to extract its features at five scales, which can compute the normalized inner distance and pixel-level classification and achieve high-throughput accurate cell segmentation (accuracy >0.9743). We validated the robustness and generalizability of our STU-Net by the accurate segmentation of data from HeLa cell slides across the full FOV and live C166 cells in vitro. Given its capability for quantifying cell growth and proliferation based on the multi-cell parameters generated from segmentation results, the proposed approach is expected to provide a strong foundation for subsequent drug development and biological screening.