| 刘珍娟,傅迎霞,张羽,彭飞,张宗军.不同CT图像重建算法下基于深度学习的肺结节检测算法效能[J].中国医学影像技术,2019,35(12):1775~1779 |
| 不同CT图像重建算法下基于深度学习的肺结节检测算法效能 |
| Effect of CT image reconstruction methods on performance of pulmonary nodules detection algorithm based on deep learning |
| 投稿时间：2019-05-06 修订日期：2019-10-29 |
| DOI：10.13929/j.1003-3289.201909048 |
| 中文关键词: 肺结节 深度学习 重建算法 体层摄影术,X线计算机 |
| 英文关键词:pulmonary nodules deep learning reconstruction algorithm tomography, X-ray computed |
| 基金项目: |
| 作者 | 单位 | E-mail | | 刘珍娟 | 南京中医药大学附属中西医结合医院放射科, 江苏省中医药研究院, 江苏 南京 210028 | | | 傅迎霞 | 南京中医药大学附属中西医结合医院放射科, 江苏省中医药研究院, 江苏 南京 210028 | | | 张羽 | 南京中医药大学附属中西医结合医院放射科, 江苏省中医药研究院, 江苏 南京 210028 | | | 彭飞 | 南京中医药大学附属中西医结合医院放射科, 江苏省中医药研究院, 江苏 南京 210028 | pengfei_82@163.com | | 张宗军 | 南京中医药大学附属中西医结合医院放射科, 江苏省中医药研究院, 江苏 南京 210028 | |
|
| 摘要点击次数: 2371 |
| 全文下载次数: 747 |
| 中文摘要: |
| 目的 探索CT图像重建算法对于基于深度学习（DL）的肺结节检测算法的影响。方法 选取298例接受肺部CT检查患者，依次采用肺窗重建、纵隔重建、骨窗重建3种算法重建CT图像。先由2名主治医师对入组病例进行标注，结果不一致时由1名高年资医师进行审核，以结果作为金标准。以深度神经网络为基础构建肺结节检测算法，与医师标注结果进行比对，得到算法在不同重建方法下检出肺结节的敏感度、准确率、F分数等指标以及模型检出的假阳性分布，对比分析模型在不同CT图像重建算法下的诊断效果。结果 基于DL的肺结节检测算法在肺重建、纵隔重建和骨重建3种重建方法下的敏感度分别为92.33%（313/339）、86.97%（287/330）及92.73%（319/344），准确率分别为23.55%（313/1 329）、37.91%（287/757）及27.84%（319/1 146），F分数分别为0.38、0.53及0.43，3种算法重建下模型检出敏感度、模型误检结节类型与医师漏标结节类型差异均无统计学意义（P均>0.05）。结论 基于DL的肺结节检测算法在肺窗、纵隔和骨窗重建下均性能优良，能帮助医生提高工作效率和诊断质量。 |
| 英文摘要: |
| Objective To explore the impact of CT image reconstruction methods on the performance of pulmonary nodule detection algorithm based on deep learning (DL). Methods Lung CT images of 298 cases were labeled by 2 attending doctors, and the inconsistent results between them were checked by a senior doctor. The final labels were regarded as the gold standards of this experiment. Pulmonary nodule detection algorithm was constructed based on a deep neural network and tested on these 298 cases. Comparing the output of the detection algorithm with the doctor's labeling, the sensitivity, accuracy and F1-score of the algorithm were calculated, especially those under different CT image reconstruction methods. Afterwards, the false-positive detections of the algorithm were checked, and the detailed distribution of these false positives was presented. Diagnostic effects of the model were analyzed among different CT image reconstruction algorithms. Results The sensitivity of pulmonary nodule detection algorithm under mediastinum, lung, and bone CT reconstruction methods was 92.33% (313/339), 86.97% (287/330) and 92.73% (319/344), while the precision was 23.55% (313/1 329), 37.91% (287/757) and 27.84% (319/1 146), respectively. Taken sensitivity and precision into account, F1-socre of these 3 reconstruction methods was 0.38, 0.53 and 0.43, respectively (all P>0.05). Conclusion Pulmonary nodule detection algorithm based on DL achieves excellent performance under pulmonary window reconstruction, mediastinum window reconstruction and bone window reconstruction, which can help doctors to improve work efficiency and diagnose quality. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
