【摘要】 目的 探究基于YOLOv5s的人工智能模型（YOLOv5s＿CBCT）对缺牙位置牙槽骨高度和宽度的测量价值，为口腔种植术前规划提供精准量化工具。方法 基于跨阶段部分融合模块（C2f）和大选择性核模块（LSK）构建C2f＿CBCT模块和LSK＿CBCT模块，并使用空间到深度下采样模块（SPD）、幽灵卷积模块（GhostConv）、幽灵C3模块（C3Ghost）和双向特征金字塔网络（BiFPN）等模块和结构，提出YOLOv5s＿CBCT模型，将测试图像输入到训练完成后的YOLOv5s＿CBCT模型中，分别得出模型对4个种植区域牙槽骨高度和宽度的预测平均值并进行误差分析；使用双向方差分析探究影响测量结果的显著因素，并进一步采用Bland-Altman一致性分析和Tukey检验来加强和完善双向方差分析的结论。结果 （1）模型对比实验表明，YOLOv5s＿CBCT模型的检测精度更高，且参数量和计算量更低。（2）模型对4个种植区域中牙槽骨高度检测的平均值误差最大为0.15 mm，最小为0.03 mm，对宽度检测的平均值误差最大为0.26 mm，最小为0.10 mm。（3）双向方差分析结果表明，牙槽骨高度和宽度的检测结果受到种植区域的显著影响，而不受测量方式的影响。（4）Bland-Altman一致性分析进一步证明了人工测量和模型预测具有较好的一致性。（5）Tukey检验确定了对测量结果产生显著影响的种植区域的分组情况。结论 YOLOv5s＿CBCT模型的测量值误差均在可接受范围内，测量结果具有较高的临床参考价值，且YOLOv5s＿CBCT模型便于在医学仪器上部署。更多还原

【Abstract】 Objective To explore the measurement value of the artificial intelligence model based on YOLOv5s(YOLOv5s＿CBCT) for the height and width of alveolar bone at the location of missing teeth, and to provide a precise quantitative tool for preoperative planning of oral implants. Methods Based on the faster implementation of CSP bottleneck with 2 convolutions(C2f), and large selective kernel(LSK) modules, the C2f＿CBCT and LSK＿CBCT modules were constructed. Using modules and structures such as space-to-depth(SPD), ghost convolution(GhostConv), C3 with ghost convolution(C3 Ghost) and bidirectional feature pyramid network(BiFPN), the YOLOv5s＿CBCT model was proposed, and the test images were input into the YOLOv5s＿CBCT model after training. The predicted average values of alveolar bone height and width for the four implant areas by the model were respectively obtained and error analysis was conducted. Bidirectional analysis of variance was used to explore the significant factors influencing the measurement results, and Bland-Altman consistency analysis and Tukey test were further adopted to strengthen and improve the conclusions of bidirectional analysis of variance. Results(1) Model comparison experiments showed that the proposed model achieved higher detection accuracy with fewer parameters and lower computational complexity.(2) The maximum average error of alveolar bone height detection in the four implant regions was 0.15 mm, and the minimum was 0.03 mm. The maximum average error of width detection was 0.26 mm, and the minimum was 0.10 mm.(3) The results of the two-way ANOVA indicated that the detection results of alveolar bone height and width were significantly affected by the implant region, but not by the measurement method.(4) Bland-Altman consistency analysis further demonstrated good agreement between manual measurements and model prediction.(5) The Tukey test identified the grouping of implant regions that had a significant impact on the measurement results. Conclusion The overall experimental results demonstrated that the measurement errors of the model were all within an acceptable range. The measurement results have high clinical reference value, and the proposed YOLOv5s＿CBCT model can be conveniently deployed on medical instruments.更多还原