<p>Accurate cell boundary segmentation in microscopy images remains challenging due to weak or low contrast boundaries, limited annotated data for training, and the inability of existing loss functions to preserve fine structural details. To address these boundary-specific limitations, we introduce a curvature-aware loss function that incorporates high-level perceptual curvature features into the optimization process. This approach formulates curvature-based regularization in a curvature embedding space that enhances the model’s sensitivity to boundary segmentation. Extensive experiments on publicly available microscopy datasets show that the proposed method improves segmentation performance, particularly for touching cells, low-contrast boundaries, even when trained with small datasets. The curvature-based loss achieves an average Dice improvement of 9.18% over Jaccard loss and up to 6.21% improvement compared with existing boundary-aware losses. These findings demonstrate that modeling boundary curvature leads to smoother, more continuous, and more accurate cell segmentations. The practical implication of this work lies in its ability to support more reliable automated cell segmentation, which is essential for downstream biomedical analysis and diagnostic applications. The originality of the method comes from explicitly encoding curvature information into the loss formulation, introducing a novel geometric regularization mechanism that enhances boundary prediction with minimal computational overhead.</p>