| 王蕾,殷秀强,龙翔,邱新,童欢珞.基于空间可分离卷积SS-3DUNet模型自动分割增强T1WI所示肛瘘瘘管[J].中国介入影像与治疗学,2024,21(11):696-701 |
| 基于空间可分离卷积SS-3DUNet模型自动分割增强T1WI所示肛瘘瘘管 |
| SS-3DUNet model based on spatially separable convolutions for automatically segmenting anal fistula on enhanced MR T1WI |
| 投稿时间：2024-09-12 修订日期：2024-09-26 |
| DOI：10.13929/j.issn.1672-8475.2024.11.010 |
| 中文关键词: 直肠瘘 磁共振成像 自动分割 深度学习 |
| 英文关键词:rectal fistula magnetic resonance imaging automatic segmentation deep learning |
| 基金项目: |
|
| 摘要点击次数: 99 |
| 全文下载次数: 38 |
| 中文摘要: |
| 目的 观察基于空间可分离卷积构建的SS-3DUNet模型自动分割增强T1WI所示肛瘘瘘管的价值。方法 回顾性分析29例肛瘘患者的2 405幅盆腔轴位增强MR T1WI，随机选取其中19例共1 537幅图像为训练集，以5例424幅图像为验证集，5例444幅图像为测试集。基于空间可分离卷积构建SS-3DUNet模型用于自动分割增强MR T1WI中的肛管瘘管，并引入层间特征强化模块加强定位瘘管特征；于训练集、验证集训练并选择最佳模型。以人工标注结果为标准，基于测试集观察SS-3DUNet模型自动分割肛瘘瘘管的效能。结果 SS-3DUNet自动分割测试集单幅图像中的瘘管用时为0.59～0.61 s，分割瘘管边界区域与人工标注区域的吻合度较高；其分割测试集瘘管的平均分割戴斯相似系数、敏感度及精确率分别为0.746、70.04%及82.93%。结论 基于空间可分离卷积SS-3DUNet能有效自动分割增强T1WI所示肛瘘瘘管。 |
| 英文摘要: |
| Objective To observe the value of SS-3DUNet model based on spatially separable convolutions for automatically segmenting anal fistula in enhanced MR T1WI. Methods Totally 2 405 pelvic axial enhanced MR T1WI of 29 patients with anal fistula were retrospectively analyzed， and 1 537 images from 19 cases were randomly selected as training set， 424 images from 5 cases were as validation set， 444 images from 5 cases were as test set. A SS-3DUNet model was constructed based on spatially separable convolutions to automatically segment anal fistula in enhanced MR T1WI， and inter-layer feature enhancement module was incorporated to improve the location of fistula features. The model was trained in training set and the best one was selected based on validation set. Taking the results of manual labeling by clinicians， the efficacy of SS-3DUNet model for automatically segmenting anal fistulas was observed based on test set. Results The time of SS-3DUNet automatically segmenting anal fistula in a single image in test set was 0.59—0.61 s， and the coincidence of the boundary of fistula segmented by the model and manual label was high. The average Dice similarity coefficient， sensitivity and accuracy of SS-3DUNet for automatically segmenting anal fistula in test set was 0.746， 70.04% and 82.93%， respectively. Conclusion SS-3DUNet model based on spatially separable convolutions could effectively automatically segmenting anal fistula in enhanced T1WI. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
| 关闭 |
