Analysis of Quantitative MRI Flashcards
Where does the Quantiative MRI aim to move from?
A qualitative picture to an objective measure
What do we want the quantitative MRI to be?
Of higher resolution, whole brain coverage and be robust to types of artefacts
What can objective measure characterise?
Tissue and can be compared between different individuals or different time points or different scanners
What is the importance of quantitate data
Quantitative data refers to numbers and statistics, and is very useful in finding patterns of behaviour or overriding themes
• R1 is 1/T1 – it is a rate
• MT – a measure of magnetisation transfer saturation – tells about the macrostructure of the tissue
• R2* is 1/RT*
What is coefficient variation?
Something that is capturing the variability when the quantity is measured multiples of times
What do you have when you have weighted imaging?
Very high coefficient variation
It is driven by hyperintense regions in the frontal areas of the brain
Sensitivity of the receiver coil high in that location
Very dependent on how the individual has positioned in the head coil
When do given voxels have arbitrary units?
When it is dependent on hardware when you look at weighted images and the actual signal intensity
What is found at every spatial location?
Measurement that is descriptor of microstructure of the water is experiencing and is a physical property of that tissue
What does the quantitative anatomical MRI (qMRI) aim to overcome?
The inter-site bias issue
It is specifically designed to provide absolute measure and thus data that are comparable across sites and time points
What is demonstrated in a qMRI>
use of quantitative mapping of the longitudinal and the transverse relaxation time (T1 and T2) in a multi-center study at 1.5T. They demonstrated a high comparability between sites and reproducibility within a single site in scan-rescan experiments (<10% deviation).
What is Multi-parameter mapping protocol?
- Axial sections
- Have distinct unit e.g. rate quantity is per second
- Tissues appear very differently depending on what quantity is being measured
What is dependent on the macrostructure?
- Different measures
2. Different types of characteristics of that tissue
What is R2* taken to be?
Marker of iron content
What does MT map measure ?
evolution of the binary spin bath after the off-resonant saturation pulse can be described by
simulataneous apparent T1 relaxation of both pools as in fast exchange and simultaneous equilibration of the partial saturation
the MT-related partial saturation of the free water can be calculated using a PD-w and a T1-w reference signal
Such
MT-sat(uration) maps are independent of the underlying T1 and largely compensated for flip angle inhomogeneities
What does MT map measure?
measures very sensitive to the many macro molecular component e.g. myelin – very strong contrast between white and grey matter, good contrast between white matter and various structures of the Basal Ganglia. This can be a very useful factor feature to exploit because weighted imaging is a combination of PD, R1, R2* and how you acquired the data, you can have different contribution from each of those
What are the characteristics of weighted images (T1-weighted image)?
- Contrast due to a combination of T1, T2, or T2*, proton density, scanner
- Arbitrary units
- Not comparable across sites or time
- Always field strength dependent
What are the characterstics of R1 map Quantitative maps?
- Contrast due to a specific MRI property
- Specific units e.g. seconds
- Comparable across sites and time
- Sometimes Field Strength Dependent, e.g. T1
What is the principle of calculating a quantiative map?
- Models of how these different data should appear and they are changed in based on particular property we want to quantify
- By applying model to the data – we can quantify what the values in the map should be at every voxel
How do you get quantitative map?
weighted images + physical model
What does R2* tell us?
The rate of decay of the transverse magnetisation specifically in a gradient echo
Calculating a quantitative map: R2*
We need a physical model to describe the signal
R2* describes the exponential signal decay over time
Our model for R2* = S0 exp(-TER2*)
We acquire data and fit to this model
What is model fitting?
We can take it as a function that we have to minimise the difference between our data and the data in the model at those particular time point
What is an example of model fitting?
Cost function which is a measure we are trying to minimise - could be the sum of squared difference between model and data
What are the steps for model fitting?
- Starting value
- calculate cost function
- Check against tolerance
- Find value
- otherwise perturb parameter
What is model fitting?
Fit by minimising or maximising some criterion: the cost function, e.g. sum of squared difference between model and data
What is model fitting: system of linear equations?
- Linearize the model
- Start off with an equation, take the natural logs of both sides
- Graph of a straight line
- The intercept is the log of S0, and we have a slope across TE but is R2*
- Therefore, find the slope of the line
- Much more efficient
- We do not need starting values and cost function
- Slope is R2*
How can we find a solution?
Linearising the model
Why do we need a biophysical model?
To describe the signal
Acquired signal is a function of flip angle, R1, TR
We acquire data and fit it to this model
How can the contrast in images be changed by?
Changing the flip angle, the degree of T1 weighting has changed between those two
What are the factors altering the apparent R1?
- Perfusion of blood into imaged voxels
- Temperature
- Imperfect spoiling of signal between TRs
- Transmit field inhomogeneity
Perfusion of blood into imaged voxels
- Unperturbed magnetisation entering the voxel will decrease the apparent T1
- Blood T1 is longer than that of tissue
Temperature
Higher temperature, longer relaxation time
What is a problem for rapid short TR sequences?
Imperfect spoiling of signal between TRs
What is transmit field inhomogeneity?
Non-uniform flip angle
Biophysical Model: R1
- The scanner hardware cannot produce it perfectly everywhere [the flip angle]
- When you have transmit field it interacts with the participant with their brain
- It can be variable by individual basis
- We can quantify it and see how inhomogeneous it is by making a calibration scan
- We can factor that calibration into our estimate by correcting the flip angle by the appropriate amount
- Flip angle map can move our points left or right
- This effect becomes more prominent as we move up in field strength
What are the two stages that registration is carried out?
Step 1 = alignment of the two images
Step 2 = re-slice the image to be moved
What does alignment of two images involve?
Varying a number of parameters to obtain a better agreement
Re-slice the image to be moved
The parameters obtained from the alignment are used and interpolation carried out on the signal intensities
How can rigid-body movement be characterised?
Three translation; x,y, and z
Three rotations: x,y,z
What is the additional factor for rigid-body movement?
- Magnification
2. Shearing
What is the process of alignment?
- First you must define a model
- Choose a number of parameters
- 6: three translations, three rotations
- 9: above + three magnification factors
- 12: above + three shears - These are rigid body registration
- assumes that the whole brain moves as one object - Minimise or maximise a cost function to determine parameters
What is Resampling?
Need to interpolate to estimate image values at new locations
- Nearest neighbour
- Tri-linear
- B-splines of varying order: higher order, higher frequency
What is another use of registration?
Create a template
Puts all subjects in same space - ‘normalised’ or ‘stereotactic’ space e.g. Montreal Nuerological institute (MNI) space
- Allows analysis of the same tissue at specific locations
What are examples of templates based controls?
- MNI 152 (T1-w)
2. FMRIMB52 (FA)
What does the template assume?
High level of correspondence across population
Can be problems with pathology
How do you segment into tissue classes?
- Separate an image into white matter, grey matter etc
- Requires a priori information: tissue probability maps
- Performance depends on priors, algorithm and contrast of image/map to be segmented
What is the caution for segmenting into tissue classes?
Pathology e.g. lesions, can cause problems since these are not expected by standard models
What is the voxel-based morphometry pipeline?
- Segment into tissue classes
- Calculate deformations
- Normalise to group space
- Modulate and spatially smooth
- Run statistical analysis
Examples of Morphometric Insight Cross-sectional analysis
- Look at taxi drivers
- How their hippocampal volume relates to their learning of knowledge to way to move around London
- Hippocampus is thought to be involved in spatial navigation
- They used automated analysis approaches
- Manually went in to define how many voxels were in the hippocampus – nice validation
- They were then able to identify areas of hippocampus that correlated with acquired spatial navigation skills
Examples of Morphometric insight: Longitudinal analysis:
- They were looking at the acquisition of the skills of juggling
- They took people of who didn’t know how to juggle and taught them how to do so for a period of months
- Acquired image data at different time points
- Stopped practising juggling and measured again months later
- A change in area associated with visual perception of complex motions
What are the sources of bias in patient population?
- Are you picking well patients?
- Is age representative of population?
- For some disease education is important - disability tests
- Confounding conditions e.g. blood pressure
What are sources of bias in control population?
- Same age and sex as patients
- Same social class/education/fitness/handedness
- Other conditions
What are sources of bias in image acquisition?
- Is your technique likely to pick up disease effect
- which parameter is appropriate?
- multi-parametric approach? - Hypothesis driven research?
- what is expected to occur?
- is there histological evidence? - Effect size?
- SNR
- Number of patients - Artefacts
What are sources of bias in image processing?
- Quality of segmentation
- Equivalent between groups - Normalisation
- smoothing carried out to account for residual misalignment after normalisation procedure
- Need to minimise partial volume effect so use a method that accounts for this e.g. VBQ - Region drawing or voxel-wise analysis?
- Appropriate statistical methods e.g. Bonferroni correction
