| 孙红,莫光萍,徐广辉,杨晨.基于融合视觉Transformer与边缘引导编码-解码网络(RET-Net)算法分割脊柱MRI[J].中国医学影像技术,2023,39(4):577~581 |
| 基于融合视觉Transformer与边缘引导编码-解码网络(RET-Net)算法分割脊柱MRI |
| Segmentation of spine MRI based on vision Transformer and edge guidance coding-decoding network (RET-Net) |
| 投稿时间：2022-11-17 修订日期：2023-03-01 |
| DOI：10.13929/j.issn.1003-3289.2023.04.021 |
| 中文关键词: 脊柱 磁共振成像 诊断,计算机辅助 |
| 英文关键词:spine magnetic resonance imaging diagnosis, computer-assisted |
| 基金项目:上海市自然科学基金项目(21ZR1450200)。 |
|
| 摘要点击次数: 1259 |
| 全文下载次数: 715 |
| 中文摘要: |
| 目的 提出融合视觉Transformer与边缘引导的编码-解码网络(RET-Net)算法,观察其分割脊柱MRI的效能。方法 基于二类分割公开脊柱数据集spinesagt2wdataset3选取195幅脊柱三维T2WI及经过标注的对应脊柱掩码,对脊柱区域与背景设置不同标签。将残差卷积网络嵌入编码-解码网络,引入边缘模块引导网络,关注脊柱边缘粒度信息并提取边缘特征;结合视觉Transformer与残差网络提取脊柱全局及局部信息,构建RET-Net分割脊柱的深度学习模型,评价其分割脊柱的效能。结果 利用RET-Net算法能准确分割脊柱椎骨区域,边缘分割较为平滑;RET-Net在数据集中的戴斯相似系数(DSC)为90.15%,交并比(IOU)为81.06%,敏感度(SE)为92.71%,特异度(SP)为99.57%,准确率(ACC)为98.61%,豪斯多夫距离(HD)为1.84 mm,其DSC及ACC等均优于UNet、PSPNet和Attention-UNet等基础分割模型。结论 融合视觉Transformer与边缘引导RET-Net算法分割脊柱MRI效能较佳。 |
| 英文摘要: |
| Objective To propose a coding-decoding network architecture (RET-Net) combined with vision Transformer and edge guidance module, and to observe its efficiency for segmentation of spine MRI. Methods Based on the public spinal dataset spinesagt2wdataset3, 195 three-dimensional T2WI of spine with corresponding labeled spinal masks were selected, and different labels were set for the spine region and the background. The residual convolution network was embedded into the coding-decoding network, and an edge module was introduced to guide the network to pay attention to the spinal edge granularity information to extract edge features. The global and local information of spine was extracted with residual network combined with Vision Transformer, then an encoding-decoding RET-Net spinal segmentation deep learning model was constructed, and its efficiency for spinal segmentation was evaluated. Results RET-Net could accurately segment the spine on MRI, and the edge segmentation was smooth. The Dice similarity coefficient (DSC), intersection over union (IOU), sensitivity (SE), specificity (SP), accuracy (ACC) and Hausdorff distance (HD) values of RET-Net in dataset was 90.15%, 81.06%, 92.71%, 99.57%, 98.61% and 1.84 mm, respectively, among which DSC and ACC were superior to those of UNet, PSPNet, Attention-UNet and other basic segmentation models. Conclusion Using RET-Net combined with vision Transformer and edge guidance module could obtain good results for segmentation of spine MRI. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
