SlideFlame: A data-efficient vision-language model for anatomically grounded pathology reporting

Abstract Pathology assessment is central to cancer diagnosis: a pathologist examines a whole-slide image and writes a report describing what they see. It remains an open question whether AI can do this end to end. Whole-slide images (WSIs) are gigapixel-scale, requiring specialised vision-language models. Existing slide-level systems have largely been trained on proprietary data and evaluated internally using simple semantic metrics, leaving generalisation and failure modes unresolved. Here, we present SlideFlame, a compact slide-level report-generating model with three contributions. First, it is data-efficient: trained on 17,336 publicly available WSIs from TCGA and GTEx, around 30× less data than the open-source state of the art, PRISM. Second, we establish rigorous evaluation on 5,899 external WSIs across seven organ systems using structured LLM-based assessment, bidirectional natural language inference and blinded multi-pathologist review. SlideFlame matched PRISM diagnostically while reducing anatomical site misalignment (9.1% vs 25.9%) and unsupported assertions (4.4% vs 18.1%); three pathologists preferred SlideFlame in 47 of 70 cases versus 13 for PRISM. Third, weights and code are publicly released, supporting reproducibility.