Prompt engineering and diagnostic accuracy of multimodal large language models in thyroid fine-needle aspiration cytology.

Role of Large language models (LLMs) in fine-needle aspiration cytology (FNAC) image analysis remain uncertain. We evaluated two LLMs - Chat GPT-4o (OpenAI) and Claude 3.5 Sonnet (Anthropic) on 63 thyroid FNAC cases, each represented by eight microscopic images (Pap and MGG, 10x/40x), using generic and structured prompts. Structured prompts improved Bethesda concordance and near-match rates but inter-rater agreement remained poor (κ ≤ 0.09). Specificity reached 100% with structured prompts, but sensitivity dropped to ≤11.8% and misclassification persisted. LLMs show potential, but domain-specific training and validation are necessary for clinical use.