【摘要】 目的 比较单阶段目标检测法YOLO-v5、YOLO-v7、YOLO-v8 3种模型对X线平片不同部位骨折线的诊断性能，选择对细小骨折线、复杂骨折线、多发骨折线精准检测的最佳模型。方法 纳入6 740幅骨折数字化X线平片影像，使用pycharm集成开发环境配置YOLO-v5、YOLO-v7、YOLO-v8目标检测模型，以端对端的方式输入影像并可视化诊断结果。使用准确率、精准率、召回率、F1值、平均精度、ROC曲线下面积等评价指标比较不同模型诊断骨折线的性能差异。由两名放射科医师使用/不使用3种诊断模型辅助，即8种诊断方式对测试集进行诊断，比较诊断时间、准确率，敏感度、特异度等差异，分析假阳性、假阴性结果的原因。结果 3种模型中YOLO-v5在综合部位组中的骨折诊断准确率最高(78.11%),四肢长骨骨折组的准确率为82.47%,脊柱骨折组的准确率为83.34%,但对骨盆和手足部位骨折的检测效能较差，准确率分别为60.15%、65.12%。YOLO-v7模型对四肢长骨、骨盆、手足骨折检测的准确率最低(30.35%、30.11%、47.68%),但脊柱骨折诊断准确率最高(88.92%)。YOLO-v8 5组骨折检测准确率分别为76.20%、81.69%、82.36%、60.52%、55.65%,性能介于其他2个模型之间。对比有无模型辅助诊断，发现YOLO-v5、YOLO-v8模型均能提升诊断医师的工作速度及诊断准确性。结论 3种模型辅助诊断效果有差异，选择适宜模型进行辅助诊断可提升医师的诊断效率和准确性。更多还原

【Abstract】 Objective To analyze the diagnostic performance of single-stage object detection models YOLO-v5, YOLO-v7 and YOLO-v8 in fracture lines across different anatomical regions on X-ray radiographs, and to determine the optimal model for precise detection of subtle, complex, and multiple fracture lines. Methods A total of 6 740 digital X-ray radiographs of fractures were included. The YOLO-v5, YOLO-v7 and YOLO-v8 models were configured in the PyCharm integrated development environment to enable end-to-end image input and visualize diagnostic results. Evaluation metrics, including accuracy, precision, recall, F1-score, average precision, and area under the ROC curve, were used to compare the models in diagnosing fracture lines. Two radiologists evaluated the test set using eight diagnostic approaches(with/without each of the three models), comparing diagnostic time, accuracy, sensitivity, specificity, and analyzing causes of false-positive results and false-negative results. Results Among the three models, YOLO-v5 achieved the highest overall diagnostic accuracy(78.11%) for comprehensive anatomical regions. YOLO-v5 demonstrated accuracies of 82.47% for fractures in the long bones of extremities, 83.34% for spinal fractures, but lower performance for pelvic and hand/foot fractures(60.15% and 65.12%, respectively). YOLO-v7 showed the lowest accuracy for long bones of extremities, pelvis, and hand/foot fractures(30.35%, 30.11%, and 47.68%) but achieved the highest accuracy for spinal fractures(88.92%). The fracture detection accuracies of YOLO-v8 for the five groups were 76.20%, 81.69%, 82.36%, 60.52%, and 55.65%, respectively, with its performance lying between the other two models. Compared with the diagnoses with or without model assistance, both YOLO-v5 and YOLO-v8 improved radiologists′ workflow speed and diagnostic accuracy. Conclusion The three models exhibited varying effectiveness in assisting diagnosis. Selecting appropriate models for specific anatomical regions can significantly enhance diagnostic efficiency and accuracy of radiologists.更多还原