| 刘珍,郭志恒,王永成,白福忠.基于DE-HR-Net模型自动测量Miniaci角[J].中国医学影像技术,2023,39(10):1560~1565 |
| 基于DE-HR-Net模型自动测量Miniaci角 |
| DE-HR-Netmodel for automatic measurement of Miniaci angle |
| 投稿时间：2023-06-30 修订日期：2023-09-01 |
| DOI：10.13929/j.issn.1003-3289.2023.10.025 |
| 中文关键词: 肌肉骨骼畸形 X线 深度学习 自动测量 |
| 英文关键词:musculoskeletal abnormalities X-rays deep learning automatic measurement |
| 基金项目:2022年度自治区卫生健康科技计划项目(202201013)、内蒙古医科大学联合项目(YKD2021LHO37)。 |
|
| 摘要点击次数: 1169 |
| 全文下载次数: 600 |
| 中文摘要: |
| 目的 建立引入密集连接、高效通道注意力机制的高分辨率网络(DE-HR-Net)模型,观察其自动测量Miniaci角的价值。方法 回顾性分析256例膝内翻患者全下肢X线片,通过数据扩增获得900幅单侧全下肢图像;基于髋关节中心点、Fujisawa点、合页点及踝关节中心点4个特征点制作数据集标签并计算Miniaci角。以DE-HR-Net构建检测特征点模型,随机选取60幅全下肢图像,观察模型自动测量与医师手动测量Miniaci角结果的差异,评价该模型测量Miniaci角的价值。结果 DE-HR-Net模型检测髋关节中心点、Fujisawa点、合页点及踝关节中心点的均方误差(MSE)分别为1.34、1.49、1.91及1.44。模型自动测量与医师所测Miniaci角分别为(12.10±3.83)°和(12.08±3.83)°,二者高度相关(r=0.991,P＜0.05)且一致性较好(ICC=0.995),平均绝对误差(MAE)为0.22°。结论 成功建立的DE-HR-Net网络模型可用于自动测量Miniaci角。 |
| 英文摘要: |
| Objective To establish a dense efficient-channel-attention high-resolution net (DE-HR-Net) model, and to observe the value of this model for automatic measuring Miniaci angle. Methods X-ray films of whole lower limbs in 256 patients with genu varum were retrospectively analyzed, and images of 900 single whole lower limbs were obtained through data augmentation. Taken hip center point, Fujisawa point, hinge point and ankle center point as feature points of measuring Miniaci angle, the label data were constructed, and Miniaci angle was calculated based on the above 4 feature points, and the feature point detection model was established based on DE-HR-Net. Then images of 60 single whole lower limbs were randomly selected. The measuring results of Miniaci angle of the model and physician were compared, and the value of DE-HR-Net model for measuring Miniaci angle was evaluated. Results The mean-squared error (MSE) of DE-HR-Net model for detecting hip center point, Fujisawa point, hinge point and ankle center point was 1.34, 1.49, 1.91 and 1.44, respectively. The Miniaci angle measured by the established model and physician was (12.10±3.83)° and (12.08±3.83)°, respectively, which were highly correlated (r=0.991, P<0.05) and with good consistency (ICC=0.995), with the mean absolute error (MAE) of 0.22°. Conclusion The successfully established DE-HR-Net model could be used to automatically measure Miniaci angle. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
