Boundary aware dual encoder learning for accurate cell segmentation

Accurate cell segmentation is a fundamental task in computational pathology and biomedical image analysis, directly impacting applications such as disease diagnosis, cancer grading, and drug discovery. Despite recent advances in deep learning, existing methods often struggle with precise boundary delineation in densely packed and morphologically complex cellular structures. In this paper, we propose a novel boundary-guided dual encoder architecture that explicitly separates semantic region representation from structural boundary modeling. The proposed framework integrates a ResNet-50 based region encoder with a lightweight ResNet-18 boundary encoder, enabling effective multi-scale feature fusion through a boundary guided attention mechanism. This design enhances segmentation accuracy, particularly along ambiguous and low contrast boundaries. Experimental results on the BCCD dataset demonstrate that the proposed model achieves a Dice score of 95.85% and pixel accuracy of 97.47%, outperforming several state-of-the-art methods while maintaining significantly lower computational complexity. Specifically, the model requires only 12.70 M parameters and 36.86 GFLOPs, making it suitable for deployment in resource-constrained medical environments.