文献速递口腔修复与种植

多模态大语言模型在全景片种植体定位中的评估

Assessing multimodal large language models for localizing dental implant fixtures on panoramic radiographs.
PubMed 文献 · J Dent · 2026-02-16

研究评估GPT-4o、OpenAI o3和GPT-5T三种多模态大语言模型在全景放射片上定位种植体的能力,使用82张有种植体图像和82张无种植体图像,每张图像固定提示运行5次。结果显示OpenAI o3和GPT-5T的微敏感度分别为68.82%和65.66%,显著优于GPT-4o的16.97%,但中等敏感度、修复体驱动错误和运行间变异性限制自主临床应用。

要点

  • 图像级别完全检测率OpenAI o3为59.02%,GPT-5T为56.59%,GPT-4o仅2.20%。
  • 阻射修复体显著降低无种植体图像的特异性,导致假阳性增加。

核心结论

  • 推理型多模态LLM在零样本种植体定位中优于GPT-4o并减少假阳性,但中等敏感度、修复体驱动错误和运行间变异性限制自主临床应用。

研究设计与样本

  • 使用开源数据集评估82张有种植体全景放射片(297个种植体)和82张无种植体图像(按有无阻射修复体各41张平衡)。
  • 测试GPT-4o、OpenAI o3和GPT-5T三种多模态LLM,每张图像固定视觉定位提示独立运行5次。
  • 采用自由响应定位框架下的任意重叠规则评分。

主要发现

  • 有种植体图像上,微敏感度GPT-4o为16.97%,OpenAI o3为68.82%,GPT-5T为65.66%。
  • 图像级别完全检测率GPT-4o为2.20%,OpenAI o3为59.02%,GPT-5T为56.59%。
  • 有种植体图像假阳性每图像GPT-4o为3.83,OpenAI o3为1.48,GPT-5T为1.52。
  • 无种植体图像特异性GPT-4o为32.68%,OpenAI o3为65.85%,GPT-5T为68.54%;假阳性每图像分别为1.95、1.04和0.92。
  • 五次运行均检测到的种植体比例GPT-4o为1.01%,OpenAI o3为22.22%,GPT-5T为25.93%。

关键数值与推荐

  • 微敏感度:GPT-4o 16.97%,OpenAI o3 68.82%,GPT-5T 65.66%。
  • 完全检测率:GPT-4o 2.20%,OpenAI o3 59.02%,GPT-5T 56.59%。
  • 无种植体图像特异性:GPT-4o 32.68%,OpenAI o3 65.85%,GPT-5T 68.54%。

临床怎么用

  • 当前多模态LLM可用于种植体相关放射工作流程的辅助定位,但敏感度中等且运行间变异性大,不可用于自主临床决策。
  • 阻射修复体会显著降低模型特异性,临床应用中需注意假阳性风险。

局限性

  • 中等敏感度、修复体驱动错误和运行间变异性限制自主临床使用。

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

OBJECTIVES: To assess whether general-purpose multimodal large language models (LLMs) can localize dental implant fixtures on panoramic radiographs and to quantify false positives on implant fixture-absent images. METHODS: Using an open-source dataset, we evaluated 82 implant fixture-present panoramic radiographs (297 fixtures) and 82 implant fixture-absent images balanced by present or absent radiopaque restorations (41 each). We tested three multimodal LLMs (GPT-4o, OpenAI o3, and GPT-5T) with a fixed visual-grounding prompt across five independent runs per image. We scored the outputs using an any-overlap rule within a free-response localization framework. The outcomes on the implant fixture-present images were fixture-level micro sensitivity, image-level complete detection rate (CDR), and false positives per image (FPPI+). The outcomes on the implant fixture-absent images were image-level specificity (no-box rate) and FPPI-. RESULTS: On the implant fixture-present images, micro sensitivity was 16.97 % for GPT-4o, 68.82 % for OpenAI o3, and 65.66 % for GPT-5T; CDRs were 2.20 %, 59.02 %, and 56.59 %; and FPPI+ values were 3.83, 1.48, and 1.52, respectively. On the implant fixture-absent images, specificity values were 32.68 %, 65.85 %, and 68.54 %, and FPPI- values were 1.95, 1.04, and 0.92, respectively. Radiopaque restorations markedly reduced specificity. The fixtures detected in all five runs were 1.01 % (GPT-4o), 22.22 % (OpenAI o3), and 25.93 % (GPT-5T). CONCLUSIONS: Reasoning-focused multimodal LLMs outperformed GPT-4o in zero-shot implant fixture localization and reduced false positives, but moderate sensitivity, restoration-driven errors, and run-to-run variability limit autonomous clinical use. CLINICAL SIGNIFICANCE: This benchmark clarifies the current capabilities and limitations of multimodal LLMs for implant-related radiographic workflows.

来源:J Dent,查看原文。版权归原作者所有。

引用

  1. Mine Y, Taji T, Takeda S, et al. Assessing multimodal large language models for localizing dental implant fixtures on panoramic radiographs. J Dent. 2026. PMID: 41707827. doi:10.1016/j.jdent.2026.106580

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。