肌骨放射学医学插图报告:三种AI模型的评估
两名肌骨放射科专科医师独立评估解剖准确性和临床实用性。结果显示Google Gemini 3.0 Pro表现最稳定,解剖准确约40-42%,临床有用约60-65%,而所有模型在复杂病例中均出现重大错误。
要点
- AI生成的肌骨影像插图解剖准确性仍不稳定,不能无监督临床使用,需放射科医师验证。
- 单一主病灶的简单病例插图准确率较高,复杂多结构或多平面病例重大错误多见。
- 三种模型中Google Gemini 3.0 Pro性能最稳定,ChatGPT与Perplexity常出现解剖不准确。
核心结论
- AI生成的医学插图(REMIL)有望作为肌骨放射学沟通辅助工具,可快速提供影像表现的视觉摘要。
- 当前AI模型解剖准确性不一致,尚不可靠用于无监督临床使用,实施时必须经放射科医师验证。
研究设计与样本
- 使用三种付费多模态AI系统:ChatGPT(GPT-4 + DALL-E 3)、Perplexity AI、Google Gemini 3.0 Pro。
- 向三种模型提供相同报告文本和标准化提示,仅基于报告描述生成代表性插图。
- 由两名肌骨放射科专科培训医师独立评估解剖准确性和临床实用性,错误分为轻微和重大两类。
- 同时记录图像生成时间。
关键影像表现/征象
- Google Gemini 3.0 Pro生成的插图解剖准确比例约为40-42%,临床有用比例约60-65%。
- ChatGPT和Perplexity AI常生成视觉上合理但存在明显解剖错误的插图。
- 所有模型在涉及多个解剖结构或多个成像平面的复杂病例中均出现重大错误。
- 单一主病灶的简单病例由所有模型插图准确率更高。
诊断标准与关键数值
- Google Gemini 3.0 Pro解剖准确率约40-42%,临床有用率约60-65%。
- ChatGPT和Perplexity AI的解剖准确率和临床有用率原文未给出具体数值。
临床应用与阅片要点
- AI生成的医学插图可作为影像报告的文字补充,帮助临床医生和患者理解影像表现。
- 日常使用时应优先用于单一主病灶的简单病例,复杂病例插图需重点核实解剖准确性。
- 所有AI生成插图在纳入报告或交付临床前必须经放射科医师逐例验证。
局限性
- 原文报告所有模型均出现重大错误,尤其复杂病例中明显。
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)。
英文原文(摘要)· 点击展开
OBJECTIVE: Radiology reports remain predominantly text-based, requiring clinicians and patients to mentally reconstruct imaging findings. Reports in Medical Illustration (REMIL) represent an emerging approach in which artificial intelligence (AI) generates simplified visual summaries directly from report text. This study aimed to evaluate the feasibility, anatomical accuracy, and clinical utility of AI-generated REMIL in musculoskeletal (MSK) radiology.
METHODS: Twenty-five MSK imaging cases were selected. Identical report text and standardized prompts were provided to three premium multimodal AI systems-ChatGPT (GPT-4 with DALL-E 3), Perplexity AI, and Google Gemini 3.0 Pro. Each model generated a representative illustration based solely on the report description. Two fellowship-trained musculoskeletal radiologists independently assessed each illustration for anatomical accuracy and clinical usefulness. Errors were categorized as minor or major, and image-generation time was recorded.
RESULTS: Google Gemini 3.0 Pro demonstrated the most consistent performance, producing anatomically accurate illustrations in approximately 40-42% of cases and clinically useful images in 60-65% of cases, whereas ChatGPT and Perplexity AI frequently generated visually plausible images with substantial anatomical inaccuracies. Major errors were observed across all models, particularly in complex cases involving multiple anatomical structures or imaging planes. Simpler cases with a single dominant abnormality were illustrated more accurately by all models.
CONCLUSION: AI-generated REMIL holds promise as an adjunctive tool for enhancing communication in MSK radiology by providing rapid visual summaries of imaging findings. However, current AI models exhibit inconsistent anatomical accuracy and are not yet reliable for unsupervised clinical use. REMIL should therefore be implemented only with radiologist validation.
引用
- Reddy, Susmitha, Botchu, et al. "Reports in Medical Illustration (REMIL) in Musculoskeletal Radiology: An Evaluation of Evolving AI Models". Acad Radiol. 2026. PMID: 42697749. doi:10.1016/j.acra.2026.08.082
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。