학술논문
基于SPECT全身骨扫描的YOLOv5x深度学习网络模型诊断良、恶性骨病灶 / YOLOv5x deep learning network model based on SPECT whole body bone scanning for diagnosing benign and malignant bone lesions
Document Type
Academic Journal
Author
Source
中国医学影像技术 / Chinese Journal of Medical Imaging Technology. 39(12):1867-1871
Subject
Language
Chinese
ISSN
1003-3289
Abstract
目的 基于SPECT全身骨扫描构建YOLOv5x深度学习网络模型,观察其诊断良、恶性骨病灶的价值.方法 纳入699例接受SPECT骨扫描患者共5 182处骨病变,包括恶性3 105处、良性2 077处.按8∶1∶1将1 121幅骨扫描图像分为训练集(n=897)、验证集(n=112)及测试集(n=112).对训练集及验证集数据进行增强后输入YOLOv5x深度学习网络进行训练得到模型,基于测试集评估模型识别良、恶性骨灶的敏感度、特异度和准确率,及其诊断结果与金标准的一致性.结果 骨扫描YOLOv5x深度学习网络模型识别恶性骨病变的敏感度为95.75％、特异度为87.87％、准确率为91.60％,识别良性骨病灶分别为91.62％、94.38％及93.14％.模型识别骨扫描图像中骨病灶的曲线下面积(AUC)为0.98,识别恶性、良性骨病灶的AUC分别为0.97、0.98.模型诊断恶性及良性骨病灶的结果与金标准的一致性均好(Kappa=0.83、0.86,P均＜0.05).结论 基于SPECT全身骨扫描建立的YOLOv5x深度学习网络模型有助于诊断良、恶性骨病灶.
Objective To construct YOLOv5x deep learning network model based on SPECT whole body bone scanning,and to observe its value for diagnosing benign and malignant bone lesions.Methods Totally 699 patient who underwent SPECT bone scanning were enrolled,with a total of 5 182 bone lesions,including 3 105 malignant and 2 077 benign lesions.Then 1 121 bone images were divided into training set(n=897),validation set(n=112)or test set(n=112)at the ratio of 8∶1∶1.After augmentation on training set and validation set,the data were inputted to YOLOv5x deep learning network for training to obtain a recognition model.The sensitivity,specificity and accuracy of this model for diagnosing benign and malignant bone lesions were analyzed,and the consistency between its diagnosis results and gold standards was evaluated based on test set.Results The sensitivity,specificity and accuracy of bone scanning YOLOv5x deep learning network model for identifying malignant bone lesions was 95.75％,87.87％and 91.60％,respectively,and for identifying benign bone lesions was 91.62％,94.38％and 93.14％,respectively.The area under the curve(AUC)of this model for identifying bone lesions on bone scanning images was 0.98,for malignant and benign bone lesions was 0.97 and 0.98,respectively.The consistency between the diagnosis results of this model for malignant and benign bone lesions and the gold standards were both good(Kappa=0.83,0.86,both P＜0.05).Conclusion YOLOv5x deep learning network model based on SPECT whole body bone scanning was helpful for diagnosing benign and malignant bone lesions.
Objective To construct YOLOv5x deep learning network model based on SPECT whole body bone scanning,and to observe its value for diagnosing benign and malignant bone lesions.Methods Totally 699 patient who underwent SPECT bone scanning were enrolled,with a total of 5 182 bone lesions,including 3 105 malignant and 2 077 benign lesions.Then 1 121 bone images were divided into training set(n=897),validation set(n=112)or test set(n=112)at the ratio of 8∶1∶1.After augmentation on training set and validation set,the data were inputted to YOLOv5x deep learning network for training to obtain a recognition model.The sensitivity,specificity and accuracy of this model for diagnosing benign and malignant bone lesions were analyzed,and the consistency between its diagnosis results and gold standards was evaluated based on test set.Results The sensitivity,specificity and accuracy of bone scanning YOLOv5x deep learning network model for identifying malignant bone lesions was 95.75％,87.87％and 91.60％,respectively,and for identifying benign bone lesions was 91.62％,94.38％and 93.14％,respectively.The area under the curve(AUC)of this model for identifying bone lesions on bone scanning images was 0.98,for malignant and benign bone lesions was 0.97 and 0.98,respectively.The consistency between the diagnosis results of this model for malignant and benign bone lesions and the gold standards were both good(Kappa=0.83,0.86,both P＜0.05).Conclusion YOLOv5x deep learning network model based on SPECT whole body bone scanning was helpful for diagnosing benign and malignant bone lesions.