Multi-resolution T1w Brain MRI simulated data

This simulated data is based on the patient-specific brain phantoms that are generated by utilizing high resolution real subject 3D brain MRI data and performing automatic segmentations for all brain tissues. The brain MRI imaging dataset is obtained from the HCP healthy young adult sample. We selected two hundred unprocessed structural T1w brain MRI scans for phantom generation. The phantoms are created by applying the Philips proprietary automated complete brain segmentation tool on HR T1w structural MRI data.

Tissue parameters including longitudinal relaxation time (T1), transverse relaxation time (T2) and proton density (PD) are assigned to each tissue in each phantom. These parameters are collected from literature and random samples are taken from the bounded Gaussian bell distribution of given mean and standard deviation for each
tissue relaxation times.

MRI data is simulated for T1w brain MRI using our Matlab-based simulation framework. Using the phantoms as input, MRI data is simulated at two different resolutions for voxel aligned paired data, i.e. HR and LR of 0.7𝑚𝑚 and 1𝑚𝑚 respectively in both phase encoding directions. A fixed slice thickness of 1𝑚𝑚 (axial) is used for both HR and LR data. To have HR-LR paired data aligned on voxel level, zero padding in k-space is performed for LR data, i.e. the size of the LR k-space is made equal to the size of the HR k-space and then Fourier reconstructions are performed. The resulting size of simulated data is 320x320x179 for both HR and LR MRI. Ernst angle solution for a gradient echo (GRE) sequence with fixed sequence parameters of TR 18𝑚𝑠, TE 10𝑚𝑠 and FA of 30◦ is used to simulate all pairs of data. The T1w intensities are computed for each voxel. A complex Gaussian noise (standard deviation 0.4) is generated, which is derived from the simulated WM (intensity range 0−4095) in the MRI volume. The generated complex Gaussian noise is added to the real and imaginary part of the k-space data before FFT reconstructions.

It is essential to emphasize that our simulation process intentionally did not introduce any MR imaging artifacts.

The folder structure of the paired HR-LR database is as follow:
1. HR (contains 200 T1w brain MRI simulated images in Nifti format at in-plane resolution of 0.7x0.7𝑚𝑚 with slice thickness of 1𝑚𝑚)
2. LR (contains 200 T1w brain MRI simulated images in Nifti format at in-plane resolution of 1x1𝑚𝑚 with slice thickness of 1𝑚𝑚)
3. Figures (the resulting figures of brain MRI SR in png format for the associated article "Effective deep-learning brain MRI super resolution using simulated training data")

本仿真数据基于通过采用高分辨率真实受试者三维脑部MRI数据进行自动分割以获取所有脑组织所生成的患者特异性脑伪影。脑部MRI成像数据集来源于HCP健康青年成人样本。为生成伪影，我们选取了二百张未经处理的T1w结构脑部MRI扫描。通过在HR T1w结构MRI数据上应用Philips专有的全自动完整脑部分割工具，创建了这些伪影。每个伪影中的每种组织被赋予了纵向弛豫时间（T1）、横向弛豫时间（T2）和质子密度（PD）等组织参数。这些参数源自文献资料，并且从每种组织弛豫时间的给定均值和标准差的Gaussian bell分布中随机抽取样本。利用基于Matlab的仿真框架，我们对T1w脑部MRI数据进行仿真。以伪影为输入，在两个不同的分辨率下对像素对齐的配对数据进行仿真，即分别在相位编码方向上实现0.7mm和1mm的HR和LR分辨率。使用固定的切片厚度1mm（轴向）对HR和LR数据进行处理。为了实现HR-LR配对数据在像素级别的对齐，对LR数据在k空间进行零填充，即将LR k空间的大小调整为与HR k空间相同，然后进行傅里叶重建。仿真数据的最终尺寸为320x320x179，适用于HR和LR MRI。使用具有固定序列参数TR 18ms、TE 10ms和FA 30°的梯度回波（GRE）序列的Ernst角度解法来模拟所有配对数据。计算每个体素的T1w强度。生成一个复高斯噪声（标准差0.4），该噪声源自MRI体积中仿真的白质（强度范围0-4095）。将生成的复高斯噪声添加到k空间数据的实部和虚部之前进行FFT重建。必须强调，我们的仿真过程有意避免了任何MR成像伪影的引入。配对HR-LR数据库的文件夹结构如下：1. HR（包含200张T1w脑部MRI仿真图像，以0.7x0.7mm的平面分辨率和1mm的切片厚度存储于Nifti格式中）2. LR（包含200张T1w脑部MRI仿真图像，以1x1mm的平面分辨率和1mm的切片厚度存储于Nifti格式中）3. 图像（与相关文章《有效使用模拟训练数据实现深度学习脑部MRI超分辨率》相关的脑部MRI SR结果图像，以png格式存储）。