文献速递脊柱脊髓外科

通用大语言模型用于脊柱影像诊断:脊柱侧弯与腰椎管狭窄检测性能的早期分析

Generalist Large-Language Models for Spine Imaging Diagnostics: An Early Analysis of Detection Performance for Scoliosis and Lumbar Stenosis.
PubMed 文献 · World Neurosurg · 2026-05-02

本研究系统评估5个主流多模态大语言模型(Grok 2、Grok 3、Grok 4、ChatGPT、Gemini)在171张脊柱正位X线片(100例脊柱侧弯、71例正常)和200张腰椎轴位T2加权MRI(100例重度狭窄、100例正常)上检测脊柱侧弯和腰椎管狭窄的性能。结果显示模型对脊柱侧弯检测准确率较高(Grok 4达0.942),但对腰椎管狭窄识别困难(最高仅Gemini 0.600)。ChatGPT在错误回答时置信度校准较好,模型迭代间性能不一致提示需专门医学影像训练。

要点

  • 多模态大语言模型检测脊柱侧弯准确率较高,Grok 4达0.942。
  • 检测腰椎管狭窄性能普遍较低,最高为Gemini 0.600。
  • ChatGPT对错误回答的置信度校准较好,提示可靠性更优。

核心结论

  • 大语言模型在脊柱侧弯检测中表现优异,但在腰椎管狭窄识别中表现不佳。
  • ChatGPT的置信度校准更优,提示其可靠性增强。
  • 模型迭代间性能不一致(如Grok 3低于Grok 2),凸显专门医学影像训练的必要性。
  • 尽管在简单脊柱病变的患者教育中有前景,但临床采用或广泛患者使用前仍需在准确性和置信度指标上取得实质性进步。

研究设计与样本

  • 评估5个多模态大语言模型(Grok 2、Grok 3、Grok 4、ChatGPT、Gemini)检测脊柱侧弯和腰椎管狭窄。
  • 使用171张全长前后位X线片(100例脊柱侧弯、71例正常)和200张腰椎轴位T2加权MRI(100例重度狭窄、100例正常),均来自公共数据库。
  • 模型无示例提示下识别病理并量化确定性(0%-100%)。
  • 分析包括McNemar检验比较准确率,ANOVA比较置信水平。

主要发现

  • 脊柱侧弯检测准确率:Grok 4为0.942,Gemini为0.912,Grok 2为0.890,ChatGPT为0.643,Grok 3为0.637。
  • 腰椎管狭窄检测准确率:Gemini为0.600,Grok 4为0.575,ChatGPT为0.545,Grok 2为0.500,Grok 3为0.450。
  • 所有模型在各类病理中平均确定性均>70%(标准差<5.3%)。
  • ChatGPT和Grok 3在脊柱侧弯错误回答中置信度降低(P < 0.0001),仅ChatGPT在狭窄错误回答中置信度降低。
  • Gemini在狭窄错误回答中报告置信度升高(P < 0.0001)。

临床怎么用

  • 当前大语言模型可用于简单脊柱病变(如脊柱侧弯)的患者教育,但不宜用于临床诊断。
  • 在临床采用或广泛患者使用前,需大幅提升模型对腰椎管狭窄等病变的准确性和置信度指标。

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

BACKGROUND: Web-based large language models (LLMs) are increasingly used by patients for medical self-assessment, but their efficacy in spine imaging diagnostics remains underexplored. This study systematically evaluated 5 leading multimodal LLMs-Grok 2, Grok 3, Grok 4, ChatGPT, and Gemini-for detecting scoliosis and lumbar spinal stenosis across radiographs and MRI modalities. METHODS: We assessed 171 full-length anterior-posterior radiographs (100 with scoliosis, 71 normal) and 200 axial T2-weighted lumbar spine MRIs (100 with severe stenosis, 100 normal) from public databases. Models were prompted without examples to identify pathology and quantify certainty (0%-100%). Analyses included McNemar's test for accuracy and ANOVA for confidence levels. RESULTS: In scoliosis detection, Grok 4 exhibited superior accuracy (0.942), followed by Gemini (0.912), Grok 2 (0.890), ChatGPT (0.643), and Grok 3 (0.637). For stenosis, Gemini performed best (0.600), then Grok 4 (0.575), ChatGPT (0.545), Grok 2 (0.500), and Grok 3 (0.450). All models sustained >70% mean certainty (SD < 5.3%) across pathologies. ChatGPT and Grok 3 demonstrated reduced confidence in erroneous scoliosis responses (P < 0.0001), while only ChatGPT did so for stenosis. Gemini reported elevated confidence in incorrect stenosis responses (P < 0.0001). CONCLUSIONS: LLMs perform highly in scoliosis detection but struggle to identify lumbar stenosis. ChatGPT's superior confidence calibration suggests enhanced reliability. Performance inconsistencies across model iterations (e.g., Grok 3 underperforming Grok 2) underscore the necessity for specialized medical imaging training. Although promising for patient education in simple spine conditions, substantial advancements in accuracy and confidence metrics are essential prior to clinical adoption or broad patient utilization.

来源:World Neurosurg,查看原文。版权归原作者所有。

引用

  1. Hoglund ZT, Wu AQ, Kathawate VG, et al. Generalist Large-Language Models for Spine Imaging Diagnostics: An Early Analysis of Detection Performance for Scoliosis and Lumbar Stenosis. World Neurosurg. 2026. PMID: 42082102. doi:10.1016/j.wneu.2026.125023

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。