| 刘高平,曲太平,许强,张其锐,李秀丽,张志强,卢光明.基于深度学习重建常规头部2D T1WI超分辨率图像质量[J].中国医学影像技术,2022,38(3):326~331 |
| 基于深度学习重建常规头部2D T1WI超分辨率图像质量 |
| Imaging quality of super-resolution reconstruction of conventional head 2D T1WI based on deep learning |
| 投稿时间：2021-06-04 修订日期：2021-10-24 |
| DOI：10.13929/j.issn.1003-3289.2022.03.002 |
| 中文关键词: 脑 人工智能 磁共振成像 质量控制 |
| 英文关键词:brain artificial intelligence magnetic resonance imaging quality control |
| 基金项目:国家自然科学基金（81790653、81871345）、国家重点研发计划（2018YFA0701703）。 |
|
| 摘要点击次数: 1943 |
| 全文下载次数: 609 |
| 中文摘要: |
| 目的 评价基于深度学习改良U-net （M U-net）模型的常规厚层2D T1WI超分辨率重建图像质量及其在脑形态学研究中的应用价值。方法 回顾性分析730例头部MRI，包含常规厚层2D T1WI及3D T1WI，按7∶3比例将其分为训练集（n=500）和测试集（n=230）。采用M U-net模型和传统插值算法对2D T1WI进行超分辨率重建；以3D T1WI为对照，计算并比较2种重建图像的峰值信噪比（PSNR）和结构相似度（SSIM）的差异。分别采用基于体素的形态学分析（VBM）和基于皮层的形态学分析（SBM）方法测量超分辨率重建图像及真实对照图像的相对灰质体积和皮层厚度，并以组内相关系数（ICC）比较测量结果的一致性。结果 相比插值重建图像，M U-net模型重建图像的PSNR （t=4.43，P<0.01）和SSIM （t=21.81，P<0.01）更高，M U-net模型重建图像与对照图像的相对灰质体积和皮层厚度的一致性均高于插值重建图像。对于M U-net重建图像，VBM的ICC优于SBM （t=13.00，P<0.01）。VBM结果显示，不同脑区间，小脑的ICC最低（0.68±0.14），而大脑皮质区域、尤其额叶（0.93±0.04）及运动区（0.94±0.02）的ICC较高。结论 改良M U-net模型能有效提高头部2D T1WI超分辨率重建图像质量，有助于进行VBM定量分析。 |
| 英文摘要: |
| Objective To evaluate the image quality and its applicability in brain morphology research of super-resolved conventional thick 2D T1WI based on deep learning modified U-net (M U-net) model. Methods Head MRI data of 730 cases were retrospectively analyzed, including both conventional thick 2D T1WI and 3D T1WI. The data were divided into training set (n=500) and test set (n=230) at the ratio of 7:3. Super-resolution reconstruction of 2D T1WI was performed using M U-net model and traditional interpolation algorithm. Taken 3D T1WI as control, the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) of two kinds of reconstructed images were calculated and compared. Voxel-based morphometry (VBM) and surface-based morphometry (SBM) were applied to measure the relative gray matter volume and cortical thickness of reconstructed images and controls, respectively, and intra-class correlation coefficient (ICC) method was used to compare the consistency of the measurements. Results Compared with interpolated reconstructed images, reconstructed images with M U-net model had higher PSNR (t=4.43, P<0.01) and SSIM (t=21.81, P<0.01). The consistency of the relative gray matter volume and cortical thickness of reconstructed images based on M U-net model were both higher than those of interpolated reconstructed images. For M U-net reconstructed images, ICC of VBM were higher than those of SBM (t=13.00,P<0.01). The results of VBM in different brain regions showed the cerebellum had the lowest ICC (0.68±0.14), whereas ICC of the cerebral cortex were higher, especially in the frontal lobe (0.93±0.04) and motor area (0.94±0.02). Conclusion M U-net model super-resolution reconstruction based on DL could efficiently improve image quality of conventional head 2D T1WI, hence being helpful for VBM quantitative analysis of 2D T1WI. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
