Improve accuracy of cell segmentation based on a cascaded network for edge enhancement

Precise segmentation of individual cells in images facilitates the exploration of complex biological questions. With the advancement of deep learning, the accuracy of cell segmentation has also achieved new breakthroughs, and these algorithms are now widely applied in biotechnology. Currently, there are some effective segmentation methods available, such as CellPose and Mesmer. However, these methods require high-quality boundary features of cell instances in images to produce satisfactory prediction results. Therefore, we propose a deep learning-based image preprocessing method (EE-Net) that enhances cell boundary information. Using these preprocessed images instead of the original ones can lead to more precise segmentation outcomes. The size of our preprocessing model is only 1/ 56th of the CellPose model, and when combined with the CellPose model into a cascaded model (EECNet), the total number of parameters is comparable to that of the CellPose model alone. The effectiveness of this method has been tested on the TissueNet dataset, where it achieved a significant improvement of 0.7% to 1.7% in average precision(AP).