Optimizing in-context exemplars for a generalist vision-language model to support region of interest selection in ovarian tumor grossing.

Pathological gross examination guides tissue sampling in heterogeneous ovarian tumors, where mixed components can complicate selection of representative regions. We evaluated whether a generalist large vision-language model can support gross image analysis through pathology-informed in-context exemplars. Using Qwen-2.5-Vision-Language-72B, we developed single-component exemplar optimization and histology-guided exemplar optimization to promote component-aware reasoning and region-level risk flagging. In a single-institution retrospective cohort, the optimized eight-shot strategy achieved 89.8% image-level accuracy on mixed ovarian tumors and improved region-level demarcation relevant to sampling. In borderline tumors with microinvasion or microinvasive carcinoma, model-identified regions encompassed pathologist-defined sampling areas, supporting risk flagging rather than stand-alone diagnosis. In rare malignant tumors (n = 21), the same strategy achieved 90.5% image-level accuracy. These findings suggest that domain-aligned exemplar design can enhance in-context reasoning for grossing-oriented medical image analysis.