| 李志建,郝以秀,胡菂菂,曾凌翔,赵英男,龙晚生,张伟聪.基于胸部X线片构建骨质疏松症自动化预测模型:多中心研究[J].中国医学影像技术,2026,42(3):403~407 |
| 基于胸部X线片构建骨质疏松症自动化预测模型:多中心研究 |
| Constructing an automated prediction model for osteoporosis based on chest radiographs: A multicenter study |
| 投稿时间：2025-08-29 修订日期：2025-11-07 |
| DOI：10.13929/j.issn.1003-3289.2026.03.017 |
| 中文关键词: 骨质疏松症 深度学习 X线 多中心研究 |
| 英文关键词:osteoporosis deep learning X-rays multicenter study |
| 基金项目: |
|
| 摘要点击次数: 68 |
| 全文下载次数: 47 |
| 中文摘要: |
| 目的 基于胸部X线片构建骨质疏松症(OP)自动化预测模型。方法 回顾性纳入3家医院共4 542例患者(OP 1 700例,非OP 2 842例)胸片数据,分别构建深度学习(DL)模型、临床模型及联合模型用于预测OP。DL模型采用ResNet-18与Transformer相结合混合架构,基于6×6网格图像块划分策略提取局部与全局特征。以受试者工作特征曲线的曲线下面积(AUC)、准确率和F1分数等指标评估模型效能,以决策曲线分析评估临床净收益,将Grad-CAM热力图用于可视化解释。结果 DL模型预测内部测试集和外部测试集1、2 OP的AUC分别为0.917、0.897及0.921,临床模型分别为0.839、0.759及0.532,联合模型分别为0.919、0.896及0.919。联合模型在各集的AUC均高于临床模型、在内部测试集高于DL模型;DL模型和联合模型的临床净收益均高于临床模型。Grad-CAM可视化显示,DL模型主要聚焦胸片中的骨性区域。结论 基于胸部X线片构建的自动化预测模型能有效预测OP且泛化能力良好。 |
| 英文摘要: |
| Objective To construct an automated prediction model for osteoporosis (OP) based on chest radiographs. Methods A total of 4 542 patients from 3 hospitals (1 700 OP and 2 842 non-OP) were retrospectively enrolled. Deep learning (DL) model, clinical model and combined model for predicting OP were constructed. DL model adopted a hybrid architecture combining ResNet-18 and Transformer, and extracted local and global features based on a 6×6 grid image block partitioning strategy. The efficacy of these models for predicting OP were evaluated according to indicators such as the area under the curve (AUC) of receiver operating characteristic curve, accuracy and F1 score, while their clinical net benefits were observed through decision curve analysis, and Grad-CAM heatmap was used for visual interpretation. Results AUC of DL model for predicting OP in internal test set and external test set 1 and 2 was 0.917, 0.897 and 0.921, respectively, of clinical mode was 0.839, 0.759 and 0.532, of combined model was 0.919, 0.896 and 0.919, respectively. AUC of combined model was higher than that of clinical model in all sets, also higher than that of DL model in internal test set. The clinical net benefits of both DL model and combined model were higher than that of clinical model. Grad-CAM visualization showed that DL model mainly focused on the bony areas in chest X-rays. Conclusion Automated prediction model based on chest radiographs could be used to effectively predict OP, which had good generalization ability. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
