| 肖霖,张笠,唐渔,黄玉瑶,王力航,何力,何志琴.基于全局与上下文双注意力U-Net网络于脊柱矢状位X线片中分割胸椎及腰椎[J].中国医学影像技术,2025,41(1):128~132 |
| 基于全局与上下文双注意力U-Net网络于脊柱矢状位X线片中分割胸椎及腰椎 |
| Global and contextual dual attention U-Net model for segmenting thoracic and lumbar vertebrae in spinal sagittal X-ray images |
| 投稿时间：2024-06-25 修订日期：2024-10-15 |
| DOI：10.13929/j.issn.1003-3289.2025.01.027 |
| 中文关键词: 胸椎 腰椎 X线 人工智能 自动分割 |
| 英文关键词:thoracic vertebrae lumbar vertebrae X-rays artificial intelligence automatic segmentation |
| 基金项目:贵州省科技计划(黔科合支撑[2022]一般264)。 |
|
| 摘要点击次数: 373 |
| 全文下载次数: 115 |
| 中文摘要: |
| 目的 观察全局与上下文双注意力U-Net网络于脊柱矢状位X线片中分割胸椎和腰椎的价值。方法 回顾性纳入600例青少年特发性脊柱侧弯患者、共600幅脊柱矢状位X线片。对图像进行预处理,以人工标注T4~T12和L1~L5作为参考标准。将全局注意力细化(GAR)模块和注意力空间金字塔池化(A-ASPP)模块添加至U-Net网络,采用5折交叉验证方法进行训练和验证,分析其分割矢状位X线片的性能,并与金字塔场景解析网络(PSPNet)、视觉几何组(VGG)-UNet及DeepLabv3+分割结果进行对比。结果 全局与上下文双注意力U-Net网络分割脊柱矢状位X线片中胸椎及腰椎的精确度为90.58%、敏感度为89.51%、戴斯相似系数为90.20%,均优于PSPNet、VGG-UNet及DeepLabv3+网络;损失函数和平均交并比曲线显示其收敛速度快,具有较好泛化能力。结论 基于全局与上下文双注意力U-Net网络可于脊柱矢状位X线片中有效分割胸椎及腰椎。 |
| 英文摘要: |
| Objective To observe the value of global and contextual dual attention U-Net model for segmenting thoracic and lumbar vertebrae in spinal sagittal X-ray images. Methods Totally 600 spinal sagittal X-ray images of 600 patients with adolescent idiopathic scoliosis were retrospectively enrolled. The images were preprocessed, and T4—T12 and L1—L5 were manually annotated as reference standards. The global attention refinement (GAR) module and attention-based atrous spatial pyramid pooling (A-ASPP) module were added to U-Net model, fivefold cross validation method was used for training and validation, and its performance for segmenting sagittal X-ray images was analyzed, and compared with pyramid scene parsing network (PSPNet), visual geometry group (VGG)-UNet and DeepLabv3+. Results The precision, sensitivity and Dice similarity coefficient of global and contextual dual attention U-Net model for segmenting thoracic and lumbar vertebrae in spinal sagittal X-ray images was 90.58%, 89.51%, and 90.20%, respectively, which were superior to PSPNet, VGG-UNet and DeepLabv3+. The loss function and mean intersection over union curves showed that it converged quickly and had good generalization ability. Conclusion The global and contextual dual attention U-Net model could effectively segment thoracic and lumbar vertebrae in spinal sagittal X-ray images. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
