| 朱昱霖,谢耀钦,梁晓坤,邓磊,张成龙,周炫汝,张怀岺.基于卷积神经网络注意力机制U-net校正CT图像中的金属伪影[J].中国医学影像技术,2022,38(5):753~757 |
| 基于卷积神经网络注意力机制U-net校正CT图像中的金属伪影 |
| Attention U-net based on convolutional neural network for correcting metal artifacts on CT images |
| 投稿时间：2021-08-04 修订日期：2022-02-24 |
| DOI：10.13929/j.issn.1003-3289.2022.05.028 |
| 中文关键词: 猪 神经网络,计算机 伪影 重建算法 体层摄影术,X线计算机 |
| 英文关键词:swine neural networks, computer artifacts reconstruction algorithm tomography, X-ray computed |
| 基金项目: |
|
| 摘要点击次数: 1631 |
| 全文下载次数: 501 |
| 中文摘要: |
| 目的 观察基于卷积神经网络(CNN)的注意力机制U-net(Attention U-net)校正CT图像金属伪影的价值。方法 选取1支猪前蹄,将直径7.5 mm的金属麻花钻头自其蹄部前表面穿至踝部,采集不同角度原始CT图像。分别采用Attention U-net、传统普通阈值金属伪影校正(MAR)、图像增强后的传统普通阈值MAR、Cycle生成对抗网络(GAN)MAR及手动分割MAR校正原始CT图像中的金属伪影;记录校正后每幅图像的像素点CT值、空间非均匀度(SNU)及伪影指数(AI),评估Attention U-net校正金属伪影的价值。结果 以Attention U-net校正后,金属伪影对CT图像的影响降低,细节和轮廓恢复,猪前蹄结构数据得以保留,并减少了二次伪影。相比校正前,校正后图像的振幅及像素点CT值更稳定。校正前、后图像的SUN分别为165.0(133.6,198.1)和27.2(14.4,38.7),AI分别为137.5(99.4,164.6)和29.1(21.1,38.7)。结论 采用基于CNN的Attention U-net算法校正CT图像中的金属伪影可降低计算复杂度、提高MAR效率,有助于恢复原始CT图像的完整性。 |
| 英文摘要: |
| Objective To observe the value of Attention U-net based on convolutional neural network (CNN) for correcting metal artifacts on CT images.Methods A pig's front hoof was selected. A metal twist drill with 7.5 mm diameter was used to penetrate from the surface of the foot to the ankle, and original CT images at different angles were collected. Attention U-net, traditional common threshold metal artifact correction (MAR), traditional common threshold MAR after image enhancement, cycle generative adversarial networks (GAN) MAR and manual segmentation MAR were used to correct metal artifact of original CT images, respectively. CT value of pixel points, spatial non-uniformity (SNU) and artifact index (AI) of each corrected image, and the value of Attention U-net for correcting metal artifacts were recorded.Results The impact of metal artifacts on CT images were reduced after correction using Attention U-net, while the details and contours disturbed by metal artifacts recovered, the structures were remained, and the secondary artifacts were suppressed. The amplitude and CT value of pixel points of images after correction were more stable than those before correction. SNU of images before and after correction was 165.0 (133.6, 198.1) and 27.2 (14.4, 38.7), respectively, while AI was 137.5 (99.4, 164.6) and 29.1 (21.1, 38.7), respectively.Conclusion Correcting metal artifacts on CT using Attention U-net based on CNN could reduce computational complexity, improve MAR efficiency and help to restore the integrity of original CT images. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
