探讨基于YOLO 11的肢体长骨骨折断端超声检测模型的临床应用价值。

前瞻性纳入2023年7月至2025年1月上海市第八人民医院206例长骨骨折患者，采集骨折断端的超声长轴切面图像，经Labelme软件标注，建立训练数据集（461张图像），按7:1:2随机划分为训练集、验证集和测试集。训练YOLO 11模型，通过精确率、召回率、F1分数、交并比>50%时的平均精度均值（mAP@50）评估模型性能，并与YOLOv8模型对比。另纳入40例受试者（骨折组与正常组各20例），采集骨折断端及正常骨超声图像构建独立临床验证集，通过敏感度、特异度、阴性预测值（NPV）、阳性预测值（PPV）综合评估模型的临床诊断效能，并分析单帧平均推理时间。

测试集数据显示，YOLO 11模型的精确率、召回率、F1分数、mAP@50分别为87.9%、85.3%、86.6%、89.2%，优于YOLOv8的82.9%、74.3%、78.4%、81.9%。临床验证集中，YOLO 11的敏感度（92.7% vs 83.9%）及NPV（91.2% vs 82.8%）优于YOLOv8，特异度（95.4% vs 98.0%）和PPV（96.2% vs 98.2%）略有下降，单帧平均推理时间较YOLOv8略增加（42.5 ms vs 36.1 ms）。

YOLO 11模型在肢体长骨骨折断端检测中可高效识别骨折断端，为院前快速筛查和临床诊断提供了有效的辅助工具。

To evaluate the clinical value of a YOLO 11-based ultrasound detection model for identifying fracture ends in limb long bones.

A total of 206 patients with long bone fractures in Shanghai Eighth People's Hospital from July 2023 to January 2025 were prospectively included. Ultrasonic long-axis sectional images of the fracture ends were collected. The images were labeled using the Labelme software to establish a training dataset (461 images), which was randomly divided into a training set, a validation set, and a test set at a ratio of 7:1:2. The YOLO 11 model was trained. The performance of the model was evaluated by precision, recall, F1 score, and the mean average precision when the intersection over union was greater than 50% (mAP@50), and it was compared with the YOLOv8 model. In addition, 40 subjects (20 with fractures and 20 normal controls) were newly included to construct an independent clinical validation set. Ultrasonic long-axis sectional images of the fracture ends and normal bones were collected. The clinical diagnostic efficacy was comprehensively evaluated by sensitivity, specificity, negative predictive value (NPV), and positive predictive value (PPV), and the average inference time per frame was analyzed.

According to the data of the test set, the indicators of the YOLO 11 model were superior to those of the YOLOv8 model. Its precision, recall, F1-score, and mAP@50 were 87.9%, 85.3%, 86.6%, and 89.2%, respectively. In the clinical validation set, the sensitivity (92.7% vs 83.9%) and NPV (91.2% vs 82.8%) of YOLO 11 were better than those of YOLOv8, while the specificity (95.4% vs 98.0%) and PPV (96.2% vs 98.2%) slightly decreased, and the average inference time per frame shows a marginal increase over YOLOv8 (42.5 ms vs 36.1 ms).

The YOLO 11 model can efficiently identify the fracture ends of long bones in the limbs, providing an effective auxiliary tool for rapid pre-hospital screening and clinical diagnosis.