| 武鹏宇,肖海波,于朋鑫,岳勇,齐守良,侯阳.光谱CT多模态深度学习模型自动识别急性肺栓塞并分割病灶[J].中国医学影像技术,2026,42(5):727~732 |
| 光谱CT多模态深度学习模型自动识别急性肺栓塞并分割病灶 |
| Spectral CT-based multimodal deep learning model for automatic identification of acute pulmonary embolism and segmentation of lesions |
| 投稿时间：2025-12-25 修订日期：2026-05-13 |
| DOI：10.13929/j.issn.1003-3289.2026.05.015 |
| 中文关键词: 肺栓塞 深度学习 体层摄影术,X线计算机 |
| 英文关键词:pulmonary embolism deep learning tomography, X-ray computed |
| 基金项目: |
|
| 摘要点击次数: 100 |
| 全文下载次数: 20 |
| 中文摘要: |
| 目的 观察基于光谱CT多模态深度学习(DL)模型自动识别急性肺栓塞(APE)并分割病灶的价值。方法 回顾性纳入105例接受光谱CT检查的APE患者并分为训练集(n=73)、验证集(n=10)及外部测试集(n=22)。采用SegConvNeXt分别基于肺动脉造影(CTPA)和碘密度图(IDM)构建单模态DL模型,以不同融合策略构建多模态DL模型。绘制受试者工作特征曲线,根据曲线下面积(AUC)及戴斯相似系数(DSC)评估模型识别APE及分割病灶的效能,以召回率评估最优模型定位栓子位置的效能。结果 深层特征融合多模态DL模型识别APE及分割病灶效能最高,AUC及DCS分别为0.951及0.749,均优于单模态及其他融合策略多模态DL模型;其定位中央型、肺叶型和外周型栓子的召回率分别为100%、95.45%和59.35%。结论 基于光谱CTPA和IDM的深层特征融合多模态DL模型可自动识别APE并分割病灶。 |
| 英文摘要: |
| Objective To observe the value of spectral CT-based multimodal deep learning (DL) model for automatic identification of acute pulmonary embolism (APE) and segmentation of lesions. Methods Totally 105 APE patients who underwent spectral CT examination were retrospectively enrolled and divided into training set (n=73), validation set (n=10) and external test set (n=22). SegConvNeXt was used to construct unimodal DL models based on CT pulmonary angiography (CTPA) and iodine density map (IDM), respectively. Multimodal DL models were further developed using different fusion strategies. The receiver operating characteristic curves were drawn, the area under the curve (AUC) and Dice similarity coefficient (DSC) were obtained to evaluate the performance of models for identification of APE and segmentation of lesions, while recall was used to assess the performance of the best model for localization of embolus. Results Multimodal DL model based on deep feature fusion achieved the best performance for identification of APE and segmentation of lesions, with AUC of 0.951 and DSC of 0.749, higher than those of unimodal models and multimodal DL models using other multimodal fusion strategies. Its recall for localization of central, lobar and peripheral emboli was 100%, 95.45% and 59.35%, respectively. Conclusion Multimodal DL model based on deep feature fusion of spectral CTPA and IDM images could automatically identify APE and segment lesions. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
