HarD-CASDE: Segmentation And Lineage Tracking of HeLa Cells

For a long time, cell segmentation and lineage tracing have been one of the focuses of biomedical research. At present, challenges exist in segmentation technology when dealing with cell boundary regions and small fragment cell images, and existing tracking methods underperform in tracking densely populated cell areas. Addressing these issues, we proposes a novel cell segmentation and lineage tracing method, termed HarD-CASDE. This approach merges the benefits of the HarDNet model and cascaded fusion encoder, effectively employing multi-scale information and integrating contextual features. It employs four types of harmonic densely connected networks (HarDBlock) as encoders and introduces a multi-resolution cascaded fusion encoder module. It improves boundary detection and feature segmentation, thereby enhancing the overall performance of the segmentation task. Furthermore, to enhance the segmentation of small targets, we use a combined loss function based on wBCE and wIOU to train the network. Experiments on the HeLa cell dataset show that our HarD-CASDE model significantly improves segmentation performance compared to models like Unet and TransUnet. Compared to four other classic tracking methods, our approach achieves a tracking accuracy of 0.9176. Moreover, the study verifies the significance of each module in cell segmentation and tracking tasks through ablation experiments.