A High-Efficiency and High-Accuracy Cellular Segmentation Scheme for Imperfect Cytoarchitecture Images

Accurate cellular segmentation is essential for cell morphology analysis and disease diagnosis. Traditional manual segmentation is prone to errors, while general segmentation algorithms based on deep learning often fail when dealing with imperfect cytoarchitecture images. This study proposed a high-efficiency and high-accuracy cellular segmentation scheme for such imperfect images. We first enhanced the cell images and then employed the Cellpose algorithm with the Cyto3 pretrained weight module as the foundational model. This scheme requires no additional training, ensuring high efficiency. Experimental results demonstrated a significant improvement in segmentation accuracy, achieving an IoU index of 0.86, ACC index of 0.98, MCC index of 0.91, and Dice of 0.93. When applied to mouse brain images, it successfully quantitatively displayed cell distribution density differences across brain regions. The scheme has great application potential and value in accurate biomedical research, such as quantitative analysis of cell distribution density in different brain regions and cellular localization.