Urine bubble analysis serves as a crucial diagnostic indicator for Heyi syndrome (a key Mongolian medical syndrome) and others. However, current diagnostic methods primarily depend on manual visual examination, which lacks robust automated image analysis methods. To bridge this gap, this paper propose a urine bubble image segmentation approach combining backbone replacement and loss function analysis to facilitate Mongolian medicine diagnosis. Methodologically, we adopt the Vision Mamba U-Net (VMUNet) as the base framework, and explore a backbone replacement strategy to enhance segmentation performance. Specifically, we substitute the original VMUNet backbone with four mainstream segmentation architectures—U \(^2\) -Net, U-Net, PSPNet, and DeepLabV3+—and systematically evaluate their performance on the urine bubble segmentation task. For each architecture, we adopt representative loss functions based on their prior relevance to medical image segmentation, including Binary Cross-Entropy (BCE), Dice loss, Tversky loss, Focal Tversky loss, and Boundary loss. This results in a comprehensive set of architecture–loss combinations, enabling a rigorous comparative analysis. Experimental results demonstrate that integrating the DeepLabV3+ architecture with the BCE–Dice loss function yields the best performance in edge restoration, small bubble detection, and background suppression. In particular, it achieves an Intersection over Union (IoU) of 72.88% and an F1 score of 84.31%. These findings validate the potential of this approach for urine bubble extraction and further contribute to expanding the application landscape of artificial intelligence (AI) in traditional and ethnic medical image analysis.