Abstract

Spatial transcriptomics enables characterization of cellular organization in intact tissue, but robust cell type annotation remains challenging due to heterogeneous expression profiles, mixed populations, and transitional states. Existing methods often enforce a single label per cluster, obscuring biologically meaningful ambiguity and producing overconfident assignments. We propose an ambiguity-aware, multi-stage framework for spatial cell-type annotation. The method combines hybrid spatial feature clustering with constrained language-model inference over curated label sets, and assigns confidence scores based on marker coverage, candidate separation, and entropy. Low-confidence clusters are selectively refined via local reclustering of ambiguous regions, while unresolved clusters are preserved as mixed rather than forcibly labeled. Applied to 10x Genomics Xenium spatial transcriptomics data from cholangiocarcinoma, the proposed refinement reduces cluster-level ambiguity from 16.1% to 2.27% and cell-level ambiguity from 18.4% to 0.86%, while improving confidence calibration. Spatial ablation confirms that topological integration resolves structural ambiguity over feature-only baselines, while constrained inference via a lightweight language model ensures scalable and biologically coherent annotations. These results highlight the importance of explicit ambiguity handling for reliable spatial annotation in heterogeneous tumors.