多模态大语言模型在甲状腺细胞学诊断中的性能评估
该研究评估了 ChatGPT 和 Gemini 对 100 张甲状腺细针穿刺(FNA)细胞学图像的诊断准确性和一致性,其中 50 例为甲状腺乳头状癌(PTC),50 例为非 PTC 病变。结果显示,基于 Bethesda 分类的平均诊断准确率分别为 26.2% 和 23.6%,基于具体诊断实体的准确率分别为 39.2% 和 20.2%。研究表明,尽管大语言模型在细胞学图像解读方面有潜力,但其在甲状腺 FNA 细胞学中的诊断性能有限。
要点
- ChatGPT 和 Gemini 在甲状腺 FNA 细胞学中的 Bethesda 分类诊断准确率仅为 26.2% 和 23.6%。
- 基于具体诊断实体的准确率分别为 39.2% 和 20.2%,髓样癌和未分化癌很少被正确识别。
- 提交五张图像和纳入 TIRADS 数据可在部分情况下改善性能指标。
核心结论
- ChatGPT 和 Gemini 在甲状腺 FNA 细胞学图像解读中展现出一定潜力,但诊断性能有限,凸显了当前大语言模型在该领域的局限性。
研究设计与样本
- 研究纳入 100 张甲状腺 FNA 细胞学图像,包括 50 例 PTC 和 50 例非 PTC 病变( follicular nodular disease、follicular neoplasm、medullary thyroid carcinoma、anaplastic thyroid carcinoma)。
- 代表性图像在 40× 放大倍数下捕获,并由专家病理学家复核确认参考诊断。
- 每张图像由 ChatGPT 和 Gemini 分别独立分析五轮。
主要发现
- 基于 Bethesda 分类,ChatGPT 和 Gemini 的平均诊断准确率分别为 26.2% 和 23.6%。
- 基于具体诊断实体,ChatGPT 和 Gemini 的平均诊断准确率分别为 39.2% 和 20.2%。
- ChatGPT 给出的「可疑恶性」和「意义不明的非典型性」诊断数量多于 Gemini。
- 提交五张图像和纳入 TIRADS 数据可在部分情况下改善性能指标。
- 髓样癌和未分化癌很少被正确识别。
关键数值与推荐
- Bethesda 分类平均诊断准确率:ChatGPT 26.2%,Gemini 23.6%。
- 具体诊断实体平均诊断准确率:ChatGPT 39.2%,Gemini 20.2%。
临床怎么用
- 当前大语言模型(ChatGPT 和 Gemini)在甲状腺 FNA 细胞学中的诊断准确率有限,尚不能替代细胞病理学家判读。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
Thyroid fine-needle aspiration (FNA) cytology is a widely used, cost-effective method for evaluating thyroid nodules. However, interpretation of thyroid FNA smears remains challenging, particularly in settings with limited access to experienced cytopathologists. Multimodal large language models (LLMs) have shown potential in static image interpretation in histopathology and may serve as supportive tools when expert review is unavailable. Nevertheless, data regarding their application in cytopathology, including thyroid cytology, remain limited. In this study, we evaluated the diagnostic accuracy and consistency of ChatGPT and Gemini in interpreting 100 thyroid FNA cytology images, comprising 50 cases of papillary thyroid carcinoma (PTC) and 50 non-PTC lesions, including follicular nodular disease, follicular neoplasm, medullary thyroid carcinoma, and anaplastic thyroid carcinoma. Representative images were captured at 40× magnification and reviewed by expert pathologists to confirm the reference diagnoses. Each image was analyzed by ChatGPT and Gemini across five separate rounds. For the class-based analysis, ChatGPT and Gemini achieved mean diagnostic accuracies of 26.2% and 23.6%, respectively, based on Bethesda categories, and 39.2% and 20.2%, respectively, for specific diagnostic entities. ChatGPT rendered a higher number of suspicious for malignancy and atypia of undetermined significance diagnoses than Gemini. Five-image submission and inclusion of TIRADS data improved performance metrics in some instances. Medullary and anaplastic thyroid carcinomas were rarely correctly identified. In conclusion, although ChatGPT and Gemini demonstrated the potential of LLMs in cytologic image interpretation and drafting descriptive reports, their diagnostic performance in thyroid FNA cytology was limited, underscoring the current constraints of LLMs in this domain.
引用
- Laohawetwanit T, Dixit S, Agarwal S, et al. Evaluating the diagnostic performance of multimodal large language models in thyroid cytology. Hum Pathol. 2026. PMID: 42144078. doi:10.1016/j.humpath.2026.106156
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。