Accurate liver segmentation in abdominal CT remains challenging because of low contrast boundaries and anatomical variability. Our work proposes UGRFormer, a Boundary-Aware and Uncertainty-Guided Refinement Transformer of 2.5D liver CT segmentation. The model consists of hierarchical Transformer encoder which models the global context, image-derived Sobel-based boundary priors which are injected to the multi-scale features. It also includes Uncertainty-Guided Refinement which rectifies ambiguous predictions through residual learning. The inter-slice continuity is represented in 2.5D input formulation at lower computational cost. Structural consistency is improved by using multi-scale decoding, deep supervision and auxiliary boundary prediction. Using patient-wise 5-fold cross-validation on 3D-IRCADb-01, UGRFormer achieved a mean Dice of 0.9355 and IoU of 0.8788, indicating that it has high volumetric overlap and better boundary delineation.