Traditional Mamba-UNet integrations employ four-stage architectures, replacing conventional five-stage UNets with VMamba blocks for global dependency modeling. Unlike Transformers, which suffer from quadratic complexity and high memory consumption in self-attention, Mamba-UNet achieves efficient global modeling through linear-complexity state space modeling. This paper proposes TriLVM-UNet, a lightweight three-stage architecture that integrates parameter-efficient VMamba blocks and enhances cross-stage feature interaction via an improved skip-attention bridge (SAB) module inspired by UltraLight VM-UNet. The model incorporates a Lightweight Vision Mamba (LVM) layer for high-resolution feature extraction, alongside multi-scale dilated convolution (MSDC) and convolutional block attention module (CBAM) for enhanced feature fusion. Evaluated on the 3D ACDC dataset against six baseline models, TriLVM-UNet achieves 98.57% accuracy. The GitHub repository is available at: https://github.com/730432ch/TriLVM-UNet.