FMRI Pre-Processing and Model-

Based Statistics
fMRI预处理理及基于模型的统计
• Brief intro to FMRI experiments and analysis (实
验和分析简介)

• FMRI pre-stats image processing (统计前图像处理理)

• Simple Single-Subject Statistics (单被试的统计)

• Multi-Level FMRI Analysis (多⽔水平fMRI分析)

• Advanced FMRI Analysis (⾼高级分析)

中⽂文翻译：王继源 孔亚卓
 FMRI Pre-Statistics
FMRI在统计之前

• Brief intro to FMRI analysis (fMRI分析简介)

FMRI pre-statistical image processing (统计前图像处理理):

• Reconstruction from k-space data (k空间数据重建)

• Motion correction (运动校正)
• Slice timing correction (时间层校正)
• Spatial filtering (空间滤波)
• Temporal filtering (时域滤波)
• Global intensity normalisation (全局强度标准化)
 FMRI Experiments 实验
Stimulus刺刺激

A B A B A
(baseline) (stimulus) (baseline) (stimulus) (baseline)

e.g. flashing
chequerboard 0 10 20 30 40 50
闪烁的棋盘 time (TRs)

• Simple paradigm design 简单范式设计:

- stimulus vs baseline 刺刺激vs基线
- constant stimulus “intensity” 恒定刺刺激强度
- constant block lengths 恒定block⻓长度
- many repetitions: ABABA 重复很多次的ABABA模式
• Need baseline (rest) condition to measure change
需要基线（静息）来测量量改变
 The Haemodynamic Response
⾎血液动⼒力力学响应
Basal State基线态 Activated State激活态
= HbO2⾎血红蛋⽩白
capillary beda⽑毛细⾎血管 capillary beda⽑毛细⾎血管
= Hbr脱氧⾎血红蛋⽩白

venules venules
arterioles动脉
arterioles动脉 静脉 静脉

Activation CMRO2 [Hbr] Field MRI

CBF CBV
change signal
leads to激活导致: 脑⾎血流 脑⾎血容 耗氧率 脱氧蛋⽩白
场变化 信号

• Field changes (perturbations) --> dephasing --> T2* effect

• BOLD-tuned MRI (T2*-weighted) is sensitive to this effect
场变化 (扰动) --> 失相位 --> T2* 效应
BOLD-调制过的MRI(T2*-加权) 对这个效应很敏敏感
 Predicted Response 预测响应
• The process can be modelled by convolving the
activity curve with a "haemodynamic response
function" or HRF. 这个过程可以⽤用活动曲线和HRF的卷积来建模

⊗ =
HRF Predicted neural activity预测的神经活动
⾎血氧动态反应⽅方程

time

Predicted response预测的响应
 FMRI Experiments: Analysis 分析
• Each voxel contains a time-varying
signal(BOLD signal)
每体素含随时间变化的信号（⾎血氧⽔水平依赖信号）

measured timeseries at
marked voxel标记体素的测
量量时序

time
 FMRI Experiments: Analysis 分析
• Each voxel contains a time-varying
signal(BOLD signal) BOLD response, %

每体素含随时间变化的信号（BOLD信号） 3
initial
post stimulus

• Model the stimulus-induced change in dip

2
overshoot undershoot
1

BOLD signal(predicted response) 0

time
stimulus
因刺刺激导致的BOLD信号改变建模（预测响应）
• Find which voxels have signals that match the model
找哪些体素的信号和模型匹配
• Good match implies activation related to stimulus
匹配度⾼高代表和刺刺激相关的激活

measured timeseries at
marked voxel
标记体素的测量量时序

predicted response预测响应
time
 Standard GLM Analysis 标准⼀一般线性模型分析

• Correlate model at each voxel separately 每个体素独⽴立相关的模型

• Measure residual noise variance

测量量残余噪⾳音⽅方差

• t-statistic = model fit / noise amplitude

t检验=模型拟合/噪⾳音幅值

• Threshold t-stats and display map

t检验阈值和显示图
Predicted Response
预测响应 (model)

Fitted
Amplitude
拟合幅值
Residual Noise
残留留噪声
 Standard GLM Analysis 标准⼀一般线性模型分析

• Correlate model at each voxel separately 每个体素独⽴立相关的模型

• Measure residual noise variance

测量量残余噪⾳音⽅方差

• t-statistic = model fit / noise amplitude

t检验=模型拟合/噪⾳音幅值

• Threshold t-stats and display map

t检验阈值和显示图
Signals of no interest (e.g. artifacts)
can affect both activation strength and
residual noise variance
⽆无关信号（如伪影）会同时影响激活强度和噪⾳音残差

Use pre-processing to reduce/eliminate

some of these effects
使⽤用预处理理可以减少或者消除这些效应
 FMRI Pre-Statistics 统计之前

FMRI pre-statistical image processing统计前图像处理理:

• Reconstruction from k-space data(k空间数据重建)

• Motion correction(运动校正)
• Slice timing correction(时间层校正)
• Spatial filtering(空间滤波)
• Temporal filtering(时域滤波)
• Global intensity normalisation(全局强度标准化)
 Image Reconstruction 图像重建
• Convert k-space data to images将k空间数据转换为图像:
• reconstruction algorithms重建算法
• Occasionally get problematic data偶尔会收到有问题的数据
• e.g. slice timing errors, RF spikes, RF interference
时间层错误，射频尖脉冲，射频⼲干扰

• Correct using custom-built initial analysis stages

使⽤用定制的初始分析阶段进⾏行行校正
• Scanner artefacts can be found by 检查伪影的⽅方式:
looking at data exploratory analysis 探索性分析(ICA),
 FMRI Pre-Statistics 统计之前

FMRI pre-statistical image processing统计前图像处理理:

• Reconstruction from k-space data(k空间数据重建)

• Motion correction(头动校正)
• Slice timing correction(时间层校正)
• Spatial filtering(空间滤波)
• Temporal filtering(时域滤波)
• Global intensity normalisation(全局强度标准化)
 Motion Correction: Why? 为什什么要运动校正

• People always move in the scanner

⼈人在磁共振扫描仪⾥里里总会移动
• Even with padding around the
head there is still some motion
即使带了了头垫还是会有移动

• Need each voxel to correspond

to a consistent anatomical point
for each point in time
需要保证每个时间点上的每个体素都与解剖点⼀一致
• Motion correction realigns to a
common reference
头动校正调整到共同参考位置
• Very important correction as small motions (e.g. 1%
of voxel size) near strong intensity boundaries may
induce a 1% signal change > BOLD
有⾼高强边界的⼩小幅运动会导致⾄至少1%BOLD信号改变，所以头动校正很重要
 Motion Correction 头动校正
= multiple registration多重配准

Select a MC target (reference) for all FMRI

volumes.在所有图像帧⾥里里选⼀一个校正参考⽬目标 Register each FMRI volume to target
separately
Can use either one original volume, mean of 把每个图像帧单独配准到⽬目标图像
several, standard space image etc.
可使⽤用⼀一个原始帧，⼏几个帧图像的平均，或者 Use rigid body (6 DOF)
⽤用刚体(6⾃自由度)
标准空间图像
Effect of Motion Correction 校正效果

Uncorrelated Motion 未校正

Stimulus Correlated Motion 刺刺激相关运动

Without MC 未校正 With MC 校正

 Motion Parameter Output 运动参数输出

Summary of total motion (relative and absolute)

相对和绝对运动总结

Relative = time point to next time point - shows jumps

相对=⼀一个时间点到下⼀一个时间点-显示跳动
Absolute = time point to reference - shows jumps & drifts
绝对=时间到参考-显示跳动&漂移
Note that large jumps are more serious than slower drifts,
especially in the relative motion plot
请注意，较⼤大的跳跃⽐比较慢的漂移更更严重，尤其是在相对运动图中
 FMRI Pre-Statistics 统计之前

FMRI pre-statistical image processing统计前图像处理理:

• Reconstruction from k-space data(k空间数据重建)

• Motion correction(运动校正)
• Slice timing correction(时间层校正)
• Spatial filtering(空间滤波)
• Temporal filtering(时域滤波)
• Global intensity normalisation(全局强度标准化)
 Slice Timing Correction 时间层校正

Almost all FMRI scanning takes each slice separately

⼏几乎所有的fMRI扫描都是独⽴立获得每个切⽚片。

Each slice is scanned at a slightly different time

每个切⽚片的扫描时间都会有轻微不不同。

Slice order can be interleaved (as shown) or sequential (up or down)

切⽚片顺序可以交错（如图所示）或顺序（向上或向下）
 Slice Timing 时间层
Without any adjustment, the model timing is
always the same
不不做调整时，模型时间总是相同的
Model

slice 10

slice 9
Slice 9 acquired 1.5 secs
after slice 10
切⽚片10在切⽚片9后1.5s采集
acquisition timing (TRs)
采集时间(TRs)
 Slice Timing 时间层
... but the timing of each slice’s data is different
但是实际数据的时间层是不不同的

Data

slice 10

slice 9
Slice 9 acquired 1.5 secs
after slice 10
切⽚片10在切⽚片9后1.5s采集 acquisition timing (TRs)
采集时间(TRs)
 Slice Timing 时间层
Can get consistency by shifting the data
可以通过移位数据来保持⼀一致
Data

slice 10

slice 9
Slice 9 acquired 1.5
secs after slice 10
切⽚片10在切⽚片9后1.5s acquisition timing (TRs)
采集 采集时间
 Slice Timing 时间层
... and then interpolating the data = slice
timing correction
之后插值数据，这就是时间层校正
Data

slice 10

slice 9
Slice 9 acquired 1.5
secs after slice 10
切⽚片10在切⽚片9后1.5s acquisition timing (TRs)
采集 采集时间
 Slice Timing 时间层
... and then interpolating the data = slice
timing correction
之后插值数据，这就是时间层校正
Data

slice 10

slice 9
Slice 9 acquired 1.5
secs after slice 10
切⽚片10在切⽚片9后1.5s acquisition timing (TRs)
采集 采集时间
 Slice Timing 时间层
The result of slice timing correction is that the data is changed
(degraded) by interpolation
时间层校正的结果是数据通过插值被改变（降低）了了
Data

slice 10

slice 9
Slice 9 acquired 1.5
secs after slice 10
切⽚片10在切⽚片9后1.5s acquisition timing (TRs)
采集 采集时间
 Slice Timing 时间层
Alternatively, can get consistency by shifting the
model
或者，可以通过调整模型来获得⼀一致性
Model

slice 10

slice 9
Slice 9 acquired 1.5
secs after slice 10
切⽚片10在切⽚片9后1.5s acquisition timing (TRs)
采集 采集时间
 Slice Timing 时间层
Shifted
One way to shift the Model
model is to use the 改变后
temporal 模型
derivative in the
GLM =
⼀一种调整模型的办
Original
法是利利⽤用GLM⾥里里的 Model
时间导数 原始模型
-
Temporal
Based on Derivative
Taylor approx:
时间导数
m(t+a) = m(t) + a.m′(t)
基于泰勒勒近似公式
 Slice Timing 时间层

Shifting the model also accounts for

variations in the HRF delay
调整模型同时解释了了HRF延迟的变异
• as the HRF is known to vary
between subjects, sessions, etc.
因为HRF随被试，session等变化

This is the recommended solution

for slice timing
这是解决时间层问题的推荐办法
 Motion Problems 头动问题

Motion correction eliminates gross motion changes but

assumes rigid-body motion - not true if slices acquired at
different times
运动校正可消除总体运动变化，但采⽤用刚体运动 - 如果在不不同时间获取切⽚片，则不不正确

Other motion artefacts persist including:

Spin-history changes, B0 (susceptibility) interactions &
Interpolation effects
其他运动伪影仍然存在，包括：⾃自旋历史变化，B0（磁化率）相互作⽤用和插值效应

Such artefacts can severely degrade functional results

Severity usually worse for stimulus-correlated motion
此类伪像会严重降低功能结果的强度，与刺刺激相关的运动通常会变得更更糟
 Motion Problems 头动问题
Motion correction eliminates gross motion changes but
assumes rigid-body motion - not true if slices acquired at
different times
Other motion artefacts persist including:
Spin-history changes, B0 (susceptibility) interactions &
Interpolation effects
Such artefacts can severely degrade functional results
Severity usually worse for stimulus-correlated motion
Some potential analysis remedies for motion artefacts include:
运动伪影的⼀一些潜在分析补救措施包括：
1. - including motion parameter regressors in GLM
2. 在GLM中包括运动参数回归
3. - removing artefacts with ICA denoising通过ICA去噪去除伪影
4. - outlier timepoint detection and exclusion (via GLM)
5. 离群时间点检测和排除（通过GLM）
6. - rejection of subjects displaying “excessive” motion
7. 拒绝显示“过度”运动的被试
No simple rule of thumb defining "too much" motion
没有简单的经验法则来定义“太多”的运动
 FMRI Pre-Statistics 统计之前

FMRI pre-statistical image processing统计前图像处理理:

• Reconstruction from k-space data(k空间数据重建)

• Motion correction(运动校正)
• Slice timing correction(时间层校正)
• Spatial filtering(空间滤波)
• Temporal filtering(时域滤波)
• Global intensity normalisation(全局强度标准化)
 Spatial Filtering 空间滤波
Why do it?为什什么做呢？
1. Increases signal to noise ratio if size of the
blurring is less than size of activation
如果模糊⼤大⼩小⼩小于激活⼤大⼩小，则增加信噪⽐比

2. Need minimum "smoothness" to use

Gaussian random field theory for
thresholding
需要最⼩小的“平滑度”以使⽤用⾼高斯随机场理理论进⾏行行阈值化

However: 然⽽而会
• Reduces small activation areas 减少⼩小激活区域
• Safest option is to do a small amount of smoothing
最安全的选择是做少量量的平滑处理理
• Alternative thresholding/stats eliminates the need for smoothing (e.g.
randomise, TFCE)
替代阈值/统计信息消除了了对平滑处理理的需求（例例如随机化，TFCE）
 Spatial Filtering: How? 如何操作？
Spatial filtering done by a 3D convolution with a Gaussian (cf. 1D
convolution with HRF for model)
通过⾼高斯3D卷积进⾏行行空间滤波（参⻅见模型，使⽤用HRF进⾏行行⼀一维卷积）

Weights
Each voxel intensity is replaced by a
0.1 0.3 0.4 0.3 0.1
weighted average of neighbouring
intensities 0.3 0.6 0.8 0.6 0.3
每个体素强度被相邻强度的加权平均值代替
0.4 0.8 1.0 0.8 0.4
A Gaussian function in 3D specifies
0.3 0.6 0.8 0.6 0.3
weightings and neighbourhood
size 0.1 0.3 0.4 0.3 0.1
3D中的⾼高斯函数指定权重和邻域⼤大⼩小
 Spatial Filtering: How? 如何操作？
Spatial filtering done by a 3D convolution with a Gaussian (cf. 1D
convolution with HRF for model)
通过⾼高斯3D卷积进⾏行行空间滤波（参⻅见模型，使⽤用HRF进⾏行行⼀一维卷积）

Weights
Each voxel intensity is replaced by a
weighted average of neighbouring 0.1 0.3 0.4 0.3 0.1
intensities
每个体素强度被相邻强度的加权平均值代替 0.3 0.6 0.8 0.6 0.3

1 0.4 0.8 1.0 0.8 0.4

0.5 0.3 0.6 0.8 0.6 0.3

0
FWHM 0.1 0.3 0.4 0.3 0.1

FWHM
Specify amount by Full Width Half Maximum (FWHM)
= distance between 0.5 values
通过全宽半⾼高指定数值=0.5倍值之间的距离
Spatial Filtering: Results at Different
FWHM 空间滤波：不不同全宽半⾼高的结果
 FMRI Pre-Statistics 统计之前

FMRI pre-statistical image processing统计前图像处理理:

• Reconstruction from k-space data(k空间数据重建)

• Motion correction(运动校正)
• Slice timing correction(时间层校正)
• Spatial filtering(空间滤波)
• Temporal filtering(时域滤波)
• Global intensity normalisation(全局强度标准化)
 Temporal Filtering: Why? 为何时域滤波
• Time series from each voxel contain
scanner-related and physiological
signals + high frequency noise Raw Signal原始数据
每个体素的时间序列列包含与扫描仪相关的信号和⽣生理理信号以及⾼高频噪声

• Scanner-related and physiological

signals (cardiac cycle, breathing
etc) can have both high and low
frequency components
扫描仪相关的和⽣生理理信号（⼼心脏周期，呼吸等）可能同时具有⾼高频
和低频成分

• These signals + noise hide activation

这些信号+噪⾳音隐藏激活

What is temporal filtering? 什什么是时域滤波？

• Removal of high frequencies, low

frequencies or both, without
removing signals of interest
去除⾼高低频噪⾳音，不不去掉感兴趣信号
Temporal Filtering: Highpass 时域滤波：⾼高通
Raw Signal原始数据 Highpass Filtered⾼高通滤波

• Removes low frequency signals, including linear trend

去掉低频信号，包括线性趋势
•Must choose cutoff frequency carefully (lower than frequencies of interest =
longer period)
必须仔细选择截⽌止频率（低于感兴趣的频率=更更⻓长的时间）
Temporal Filtering: Lowpass 时域滤波：低通
Raw Signal原始数据 Lowpass Filtered低通滤波

Removes high frequency noise 去掉⾼高频噪⾳音

Only useful if the predicted model does not also contain

high frequencies…
仅当预测的模型也不不包含⾼高频时才有⽤用...
 Filtering & Temporal Autocorrelation 滤波&时域⾃自相关
Some designs also contain high frequency content in the model e.g. Dense Single-Event Model:
某些设计在模型中还包含⾼高频内容，例例如密集单事件模型
Spectrum of model模型谱

• In these cases, lowpass filtering removes too much signal

这些例例⼦子⾥里里，低通去掉了了太多信号

• Also, need noise data to correctly estimate autocorrelation (to make statistics valid - see later)
➜ avoid lowpass filtering
另外，需要噪声数据来正确估计⾃自相关（以使统计数据有效-请参阅下⽂文）➜避免低通滤波

Recommendations建议:
- Use Highpass only 仅使⽤用⾼高通
- Ensure cutoff frequency higher than model frequencies (can use the Estimate button in the GUI
- see practical)
确保截⽌止频率⾼高于模型频率 （可以使⽤用GUI中的“估计”按钮-参看练习）
- Lower limit on cutoff frequency for good autocorrelation estimation(e.g. for TR=3s, cutoff
period > 90s )
较低的截⽌止频率限制可实现良好的⾃自相关估算（例例如，对于TR = 3s，截⽌止时间> 90s）
Effect of Temporal Filtering 时域滤波的影响

No Temporal Filtering⽆无时域滤波

Highpass Temporal Filtering⾼高通滤波

 FMRI Pre-Statistics 统计之前

FMRI pre-statistical image processing统计前图像处理理:

• Reconstruction from k-space data(k空间数据重建)

• Motion correction(运动校正)
• Slice timing correction(时间层校正)
• Spatial filtering(空间滤波)
• Temporal filtering(时域滤波)
• Global intensity normalisation(全局强度标准化)
 Global Intensity Normalisation 全局强度标准

• Mean intensity of the whole dataset changes

between subjects and sessions
整个数据集的平均强度会在被试和实验阶段之间变化 Subject 1

(over 3D volume)
• due to various uninteresting factors (e.g. caffeine

Mean Intensity
levels) 因为⼀一些不不相关的因素（如咖啡因⽔水平）

• Want the same mean signal level for each subject

(taken over all voxels and all timepoints: i.e. 4D)
希望每个受试者具有相同的平均信号⽔水平（在所有体素和所有时间点上
都如此：即4D）
Subject 2
Subject 3
• Scale each 4D dataset by a single value to get the
overall 4D mean (dotted line) to be the same
通过单个值来缩放每个4D数据集来使得总体4D平均值（虚线）相同
Time (TR)
• Automatically done within FEAT
在FEAT⾥里里会⾃自动完成
 Summary 总结
Reconstruction Create image and remove gross artefacts
重建 创建图像并删除严重伪影

Motion Correction Get consistent anatomical coordinates (always do

this)
运动矫正
获取⼀一致的解剖坐标（总是这样做）
Slice Timing Get consistent acquisition timing (use temporal
derivative instead)
时间层校正
获取⼀一致的采集时间（改为使⽤用时间导数）
Spatial Smoothing Improve SNR & validate GRF
空间滤波 改善SNR并验证GRF
Temporal Filtering Highpass: Remove slow drifts⾼高通：去掉低频扰动
时域滤波 Lowpass: Avoid for autocorr est.低通：避免⾃自相关

Intensity Normalisation 4D: Keeps overall signal mean

constant across sessions
强度标准化
4D：所有session期间保持整体信号平均值不不变