fMRI Principles and Data Analysis

fMRI and Standard Data Analysis

• fMRI (functional Magnetic Resonance Imaging) is a technique that measures and maps brain activity
• fMRI data consists of time series of image volumes with typical dimensions of 64x64 voxels within a slice, about 30 slices, and 150-300 volume images (time points)
• Activation during a specific task, response to task A being larger than task B, and spatial maps are typical questions for an fMRI experiment also known as "Where" questions
• The method relies on Blood-Oxygenation-Level-Dependent (BOLD) contrast

Functional Contrast in fMRI Data

• BOLD contrast relies on how oxygenated and deoxygenated hemoglobin behave in a magnetic field
• Oxygen is transported in the blood through hemoglobin
• Oxygenated hemoglobin is diamagnetic and has minimal effect on the magnetic field
• Deoxygenated hemoglobin is paramagnetic which causes small changes in the local magnetic field, resulting in faster dephasing and a decreased MR signal

fMRI Response

• fMRI signal changes with blood oxygenation and blood flow, making fMRI an indirect measure of neural activity
• Absolute magnitude of an fMRI response is typically not very useful due to its dependency on multiple factors such as voxel size, local vascular density etc
• fMRI is a relative measure, where the response strength is compared to a control/baseline condition.

Spatiotemporal Resolution of fMRI

• Specificity refers to the ability to localize events of neural origin
• Spatial resolution refers to ability to spatially separate neural events/networks
• A typical voxel size in fMRI is (2-3 mm)³
  • Smaller voxels reduce SNR and increase aquisition time
  • Larger voxels increase partiality volume effects
• Spatial resolution is good compared to other neuroimaging

Improving Spatial Resolution

• Neurovascular coupling sets the ultimate limit for spatial resolution.
• Evidence of blood flow regulation at the level of capillaries enables sub-millimeter resolution
• Spatial specificity can be increased by using an MR pulse sequence sensitive to large vessels
• Improved spatial resolution and specificity can be achieved at higher field strengths (3T->7T)

Temporal Resolution of fMRI

• Repetition time (TR) refers to the time needed to acquire one volume image, typically around 2 seconds
• Advanced MR sequences allow collection of fMRI data at TR < 300 ms
• Temporal resolution is limited by the sluggish hemodynamic response
• Poor temporal resolution is the main limitation of fMRI

Preprocessing fMRI Data

• The main goal of preprocessing is to reduce non-task-related variability in the data
• Typical preprocessing steps include data format conversion, slice timing correction, movement correction, distortion correct, spatial smoothing and spatial normalization to a brain atlas

Statistical Analysis of fMRI Data

• A typical question is determining which brain regions are activated stimulus through the task
• Standard approach includes constructing a model, fitting themodel to data, then perfoming statistical tests for each voxel

Standard Analysis using General Linear Model (GLM)

• y = XB + ε
• A "Boxcar" model shows alternating activation & rest
• Take into account shpe of the BOLD response; convolve the stimulus timing vector with a model of the hemodynamic response function (hrf)

Experimental Designs

• Blocked designs entail multiple repetitions of stimuli in blocks and have good detection power
• Event-related designs involve more stimulus types, transient activity, and good estimation power

Standard Analysis using General Linear Model (GLM) (cont.)

• Find parameters that best explain the data by minimizing the sum of the squared error values
• fMRI data fits data from one voxel one time at a time
• The t-statistic is calculated for a contrast for each voxel by dividing ( c^T \beta ) by std(( c^T \beta ))

Statistical Inference and Multiple Comparisons

• Voxel-by-voxel hypothesis testing determines if a model explains variance in the data
• Specify contrast representing the linear combination of parameter estimates
• GLM is applied independently to many voxels (100,000+) which is a "massively univariate" approach
• Multiple comparison correction correct for the possibility of chance findings
  • At a 5% level the p-value might suggest 5000 significant

fMRI as a Mapping Tool

• Involvement maps brain regions active during stimuli processing or task performance
• Specialization is where the brain contains distinct regions specialized for particular perceptual functions (e.g., faces)

Functional Specialization

• Before neuroimaging, focal lesions pinpointed specific perceptual problems Behavioral relevance for stimulus categories or tasks
• Typically always a network of brain regions
• Question: Where in Brain stimulus X Evokes a Larger Response than Stimulus Y?
• Specialized regions for face perception: fusiform face area (FFA; Kanwisher et al. 1997), occiptal face area (OFA, Gauthier 2000)

Methods of Visualization

• Volume-based visualization vs. surface-based visualization.
  • Surface-based offers better inter-subject averaging
• Brain structures are segmented from an anatomical MRI
• Cortical surface is reconstructed based on border between gray and white matter
• Reconstructed cortex is inflated into a smooth 3D surface and/or unfolded onto a 2D sheet

Visual Field Maps

• Spatial arrangement of visual field maintained in visual cortex
• Preserves arrangement with visual field maps that are retinotopic maps identified by comparing eccentricity and polar angle
• Continuum of decreasing retinotopy and increasing receptive field sizes in higher-level areas

Region-of-Interest Based Analysis

• It involves selecting a cluster of voxels or brain region a priori when investigating a region for effects
• In practice:
  • Collect Beta values for individual voxel in ROI
  • Calculate % Signal Changes for individual Voxels
  • Average across voxels in ROI
  • Average across subjects
  • Plot, show standatd-error of the mean
• Needs good justification: what are benefits and risks in ROI analysis?

Why and Risks of ROI-based Analysis

• Allows to explore data because it offers average response time
• Can be implemented to limit number of statistical tests with multiple testing
• Good to investigate function of region in detail
• One can see the Looking at the ""wrong"" region.
• Key is to avoid ""double dipping"", ""circular analysis"" to select ROIs and research question
• Kriegeskorte et al: circular analysis

Anatomical and Functional ROIs

• Anatomic ROIs must Define region-of-interest (ROI) based
• Anatomical relationship between f unction and anatomy?
• tools can be automated for segmentations in FreeSurfer
• Functional ROI involve identifying subject ROIs with independet localizers
• One combinea result across subjects and can test hypothesis with these fROIs

Functional ROIs in Visual Cortex

• retinotipc map is useful for Localizers in vision research
• Need clear functional identification criteria based on areas

Pattern-Information Analysis

• No difference: Fine-grained Information on Stim or Task Representations in Response Patterns
• Standard fMRI analysis is univariate with separate models test to the voxel
• Multivariate analysis:Multiple Voxels are tested jointly for Differences between Experimental conditions:

MVPA

• Increased sensitivity to fine-grained spatial information
• This can be achieved through distributed representations and overlapping activity patterns
• Linear multivariate is the most common
• In brain: multivocel pattern is useful for brain reading

Distributed Activity Patterns

• Can work even when Maximally Responsive VOXels are excluded.
• Can help distinguish distributed and overlapping representations.

Classifying FMRI Data

• Algorithm of Machine Learning is essential to apply to data patterns
• Key question: Can we Classify Conditions better than tasks stimuli?
• Can show activation maps to classify brain data

Classification Analysis Key Points

• Needs training data with set
• classifier can help learn training database
• Test database allows to est with reliability of stimuls patterns
• Examples: linear discrimant algorigthms

Basic Process of MVPA

• Data splitting with independent set of trinaing set
• Cross Validation is key for efficienct data use
• GLM parameter Estimates show beta T-vlaues in comparion research

VM Selectors

• Select available brain voxels to a calssifer
• Region of interest is crucial
• Searchlight approach can be adopted to study the sphere of voxels at each location of brain.

Learning with Classifier Datasets

• Can identify maximized with difference
• Can have weighting for independent datasets in learning
• Check classifier performance to determine corrext claissification

Decoding Results Key Facts

• Not mind Reading
• Focuses on Stimultask information
• Sensitivitiy to fine Grained infomration
• Lower Differences are more attenational

Data Sets for Research

• linear Classification are sensory sensitive
• But mainly have LImitations can be decodes

Representational Similarity Analysis (RSA)

• Explores perceptions
• Tests to find out whether can stimuli discriminate
• What is eye gaze directions

Relating Brain Behaviour to Model

• dismilartitiy - correlation as key concept
• Can interpret as distnances.
• Data can also directyl compare between models

More Insight to RDM

• rich Design is essntital for benefits to test group

Test Relatedness

• A matrix of dissimilarities exists between each pair of stimuli tasks
• This is determined by the calculation of a typical similarity from the correlation distance

vRF Modeling

• This uses Stimulus information as an predictor of brain activity
• It is a usefil form of "model"
• It is useful to knwo Receptive Field model too?

Usefulness for fMRI techniques

• This can enable test hypothesis
• Can model Voxel to analyze
• Is that model good for dataset.
• vRF modeling can use brain with compuataiotnal brain

Best Used Cases

• vRF helps determine when to use classification
• Need to know what best way for brain ROI to compare.
• Need to find Model for better explanable in noise

Models to Compare

• Is there specific ROI Features
• brain Activation data to make relations and predictions.
• A good compute model gives insight to brain's computations.
  • These must transform knowleged into neural representations

Explore the principles of functional Magnetic Resonance Imaging (fMRI), focusing on BOLD contrast and data analysis. Understand the role of oxygen transport, voxel dimensions, and time series analysis. Learn how fMRI experiments compare brain activity during different tasks to derive meaningful interpretations.