Abstract: To address the issues of insufficient robustness and poor physical interpretability in deep learning methods for noisy image segmentation tasks, this study proposes a novel image segmentation network called Variational Model with Mean Curvature Based Tailored UNet (VMMC-TUNet). The model innovatively integrates a physically meaningful mean curvature phase-field model with the UNet architecture, combining the Cahn-Hilliard equation with a mean curvature term and the classical UNet structure to construct an end-to-end hybrid model. In implementation, the study employs control variables and neural network approximation strategies, along with a tailored finite-point method for numerical solution. Experimental results on test datasets with various types and intensities of noise, including Gaussian noise, salt-and-pepper noise, and Poisson noise, demonstrate that VMMC-TUNet exhibits significant advantages in segmentation accuracy and edge preservation compared to existing methods, effectively enhancing the model's robustness and stability under noise interference.