Please enable JavaScript.
Coggle requires JavaScript to display documents.
Neuroimaging and psychosis - Coggle Diagram
Neuroimaging and psychosis
Structural neuroimaging
Voxel-based morphometry
Methodological limitations
Structural changes in remission may be from medication or persistent psychosis-related changes
It's all correlation, medication is a major confounder
Most studies use medicated patients
What is it?
Most common technique for analysing sMRI data
Look for significant differences in white and grey matter
Between groups
As a function of a variable, e.g. years of disease, test score
80,000-100,000 voxels in a brain image
Main steps
2) Normalization of images to a template
Registering to standard template in 3D stereotactic space
Use 1 of 2 commonly used spaces, e.g. MNI
Use a space relevant to the population sampled from
Use same space for groups you're comparing
Normalised to general shape and size, don't distort subject of interest
3) Segmentation of normalized images
into grey and white matter and CSF
use
a priori knowledge of spatial distribution of different tissues in controls
cluster analysis for voxel intensities matching different tissue types
4) Smoothing of segmented images
Because we can't be sure which voxel a signal comes from
-> more normal distribution of the signal
5) Statistical analysis of pre-processed images
t and F tests for significant effects in every voxel
correct for multiple comparisons!
beware of old studies where they haven't
e.g. false discovery rate
comparison between groups or correlation with variable of interest
1) Acquisition of neuroanatomical images using sMRI
Spatial resolution up to 1.5 mm
use the same scanner and settings within an investigation
Standardisation
Do site corrections if multicentre study
Make sure this has been done when reading papers
Subjects need to stay still
Use processing to remove regular movements but set threshold for rejection if too much movement
Takes 8-12 minutes
Main applications in early psychosis
Excluding organic causes of symptoms
Most common use
But fairly rare (c. 2% of people presenting with psychosis-like symptoms)
e.g. auto-immune encephalitis
Very different treatment, so differential diagnosis important
Developing diagnostic biomarkers
Differences between UHR and healthy controls?
e.g. Chung et al. (2018): age prediction model showed differences between CHR and HCs
Chung et al. 2019 showed where this was
Differences between FEP and healthy controls?
Differences between patients with different disorders?
Development of prognostic biomarkers
Differences between those that develop psychosis and don't?
Pantelis et al. (2003)
those that subsequently transitioned had smaller grey matter volume in
a right lateral temporal region - superior temporal gyrus, temporal region
a right inferior frontal region - inferior frontal gyrus, parts of insular and basal ganglia
a cingulate region - anterior and posterior cingular gyrus (bilateral)
a right medial temporal region - hippocampus, parahippocampal cortex
unmedicated sample
those that transitioned showed further reduction of grey matter in several areas by 12 months, while those that didn't showed only changes in the cerebellum
compared baseline scans of transitioned/didn't
also looked for longitudinal differences
Conclusions
There are some grey matter abnormalities before transition
Other appear at FEP
Mechelli et al. (2011)
Multisite study
Reduction in volume of parahippocampus in transitioner vs non-
Makes sense since parahippocampal activity drives subcortical dopamine dysfunction in animal models
So reduction in this region may be a biomarker, even if we don't know why
Tavares et al. 2022 - sMRI didn't predict transition from ARMs with machine learning approach
Collins et al. (2023): cortical thinning precedes onset and can differentiate those who will and won't convert (Lect 1)
Koutsouleris et al. (2009): machine learning with sMRI data to predict transition - 82% accuracy.
Differences between people with FEP who have multiple relapses and don't?
Development of predictive biomarkers
Differences between UHR who benefit/don't from specific treatment?
Differences between FEP that benefit/don't from specific treatment?
Reis Marques et al. (2014)
lower fractional anisotropy in non-responders than both responders and healthy control subjects in the uncinate, cingulum and corpus callosum
variety of antipsychotics used
potential biomarker
Premkumar et al. (2009): regional differences in grey matter volume in those responsive to CBT.
Great, so people don't have to go through multiple treatments and e.g. go straight to clozapine if they'll be treatment resistant
But! Neuroanatomical plasticity
Brain can change due to
Learning
Mechelli et al. (2004)
Higher grey matter density in areas associated with language acquisition in bilinguals, associated with early or late language acquisition
Maguire et al. (2000)
Higher volume posterior hippocampus in London cab drivers - experience-dependent
Gaser & Schlaug (2003)
Higher grey matter volume in auditory processing and bodily co-ordination areas in musicians, more so if have been playing longer
Draganski et al. (2006)
Studying for a university exam changed brain structure
Stress
Papagni et al. (2010)
GMV in PTSD-associated regions reduced over time in people exposed to more SLEs
Relevant to psychosis but really?!?
Merrit et al. (2023)
People with greater pre/perinatal risk had smaller left subgenual cingulate volume
Relevant to psychosis, but cross-sectional
Disease Process
Personality
IQ
Physical exercise
Other factors
Could collect data on e.g. stress to eliminate confounding, but hard to measure/control for everything
Physiological basis of MRI
sMRI: get info about neuroanatomy using signals from water in brain
Water = 2/3 of human body
Hydrogen atoms in water have spin angular momentum
Then hit them with a radiofrequency wave of that frequency
Absorb the energy
Many rotate their alignment away from the field direction
They come back to original direction by remitting energy
Measure this energy
Turn that info into voxels based on water content
Build 3-D image of the brain since different parts have different water concentrations
Molecular environment of water is altered in psychiatric/neurological disorders, so useful clinical tool
Put them in the MRI (magnet), and they align to its field and change precessional frequency
Functional neuroimaging
Physiological basis
Most common
fMRI
2 - neuronal activity also increases metabolic rate of oxygen consumption
3 - much smaller than the increase in blood flow, so local deoxhemoglobin concentration decreases below baseline
decrease is what fMRI measures
PET
Measures blood flow, metabolism, neurotransmitters or radiolabelled drugs
Measuring rate of consumption of glucose is common – through accumulation of a radiolabelled analogue of glucose
1 - presentation of stimulus, increasing brain activity
2 - elicits electrical signals from nerves to arterioles and synthesis of nitric oxide
3 - leads to muscle relaxation around microvessels (nitric oxide = informational, tells them to do it)
4 - leads to increase in cerebral blood flow (in few 100 ms)
increase in cerebral blood flow is measured by PET
neither is a direct measure of brain activity (e.g. EEG), and there's a delay
the signals are generated at different times by different processes, so comparison of PET and fMRI results is problematic
Comparison
PET
Radiation, 30-s temporal res, 2-3 mm spatial res, blocked trials only, continuous sensitivity in space, more directly related to neural activity
fMRI
No radiation, 2-3s temporal res, 2-3mm spatial res, blocked or randomised tasks, less sensitive in some brain regions, less directly related to neural activity
Do task while in scanner
Compare brain activation with two tasks, e.g. looking at 2 different types of pictures
or resting state
Functional localisation vs functional integration
Functional localisation
Different areas do different things
Brain imaging studies usually based on this principle
e.g. disrupted regional activation for fearful vs neutral faces in psychosis
identify areas specialised to tasks by comparing activity e.g. when doing task and at rest
Range of software for it
e.g. statistical parametric mapping
get a time-image series
do processing
make parameter estimates
make a statical parametric map
involves linking the stimulus presentation to the fMRI response, which isn't foolproof
Some merit - some tasks are quite localised
Functional integration
different regions communicate with each other to do tasks
alternative approaches have been developed to examine dynamic interaction
Consider the brain as a network, not a modular system
Types of connectivity
Functional
temporal correlations
Estimate by measuring correlation between regions over time
e.g. O'Neill et al. 2019 showing difference in FC in FEP
Effective
dynamic influence
Compute using complex models, e.g. Structural Equation Modelling, Dynamic Causal Modelling
e.g. disrupted effective connectivity between left superior temporal gyrus and anterior cingulate in psychosis (resting state)
Friston (2002): not mutually exclusive - functional role of a component defined by its connections
Main applications in early psychosis
Development of “diagnostic” biomarkers
Some resting-state differences in CHR, e.g. Wotruba et al. (2014)
e.g. O'Neill et al. 2019 showing difference in FC in FEP
Development of “prognostic” biomarkers
Andreou & Borgwardt (2020) review - transition
Very few task-based studies showing functional diffs in transitioners
Only non-replicated Allen et al. (2012)
Poss also in resting state (Anticevic et al. 2015)
Development of “predictive” biomarkers
Response to antipsychotics: Kapur et al. (2000) PET, Demjaha et al. (2012), Egerton et al. (2018)
OPTiMISE aims to address problem that often medicated samples
As in structural neuroimaging, but looking at differences in function rather than differences in structure
fMRI disruptions are usually during acute psychosis, subside afterwards
Barriers to translation
Studies report on the group level, but clinician has to make decisions about individuals
How does the group signal relate to the individual?
Need highly specialised methodological expertise that isn't present in clinics
Any diagnostic tool should have a flexible interface that non-specialist clinicians can use
Cost: PET and diffusion tensor imaging are very expensive
Expensive to train people and to have MRI
Research focuses on statistical significance, not clinical utility
Mismatch in incentives
Regulations
Neuroimaging packages explicitly advise against use in clinic
Lots of documentation/validation necessary if it's to get through regulatory approval for use in clinics
Note these issues are from Kempton & McGuire (2015) so maybe better now?
McGuire et al. (2015)
Something simpler like a blood test that acts as a proxy for neuroimaging findings would be ideal - need peripheral biomarkers!
Clinician attitudes: some clinicians don't like the idea because they don't subscribe to the biological model of mental illness
Methodological limitations
Data can be noisy due to
Movement-related artefacts
Individual variability in behavioural performance
Individual variability in cognition
Individual variability in functional neuroanatomy
Habituation effects at cognitive level
Changes with repetition of same task for 20 mins
Susceptibility artifacts
Where more nervous people move more
Many exclusions, e.g. having metallic implants or braces, being pregnant, not being able or willing to do the task, had PET recently
Translation: where do we go from here?
Machine learning
Supervised/unsupervised
Methods include Support Vector Machine and Cox Regression
Train the algorithm to develop a pattern classifier, then test that on a new patient - did it correctly classify?
e.g.s of use
Lalousis et al. (2012)
are clusters identified transdiagnostic for psychosis and depression?
yes - impaired group vs non-impaired, both diagnoses
developed algorithm to detect whether they'd worsened - clinical utility
Used supervised and semi-supervised ML
Koutsouleris et al. (2009)
Pattern classification for sMRI of ARMS
Scanned then followed longitudinally to detect transition
algorithm trained to discriminate those that would and would not transition at 82% accuracy
Limitations
Cross-validation - overfitting risk (Kempton & McGuire, 2015)
Help-seeking individuals - generalise to general screening?
Generalise to other equipment - needs replication
Not just based on regional differences but on covariance between regions (so it captures functional integration)
Challenges
Maximum accuracy currently at 90% - not good enough for the clinic
Training stage relies on human-made diagnoses
Would be more useful to see how good they are at classifying people who are clinically hard to classify
Ways to find biomarkers
Multicentre studies
PSYSCAN
Aim = personalised intervention - give people what will work best for them
Multicentre, multivariate, longitudinal
Multivariate machine learning to -> individualised predictions based on sMRI, fMRI, cognitive, clinical and sociodemographic data
Challenges
More heterogenous acquisition and patient population
Variable data quality
May need to develop standard MRI sequence to increase standardisation
Strengths
Increase sample size, increased power
Bring together expertise
Sample more representative
Longitudinal studies
Strengths
More power and sensitivity because person = own control
Potentially identify biomarkers signalling onset
Can unpick whether changes are due to meds, illness, substance abuse
can provide better predictors than baseline e.g. Cahn et al. (2006)
Challenges
Attrition - up to 50% reported
Scanner upgrades
Changes to research staff - differ in competence or position patients differently
Less of a problem if no differences between patient and control group
Greater logistical demands
e.g. Chung et al. (2018): Longitudinal machine-learning study showing brain-age gap different in converters/non-converters [issues!], and Chung et al. (2019) associated with surface area reductions
Multimodal imaging
Using more than one modality in each individual
e.g. Radua et al (2012) developed multimodal voxel-wise meta-analysis technique to compare results of sMRI and fMRI studies of patients with FEP
But Pettersson-Yeo et al. (2014)
in high-risk individuals found combining structural, functional and diffusion-tensor data in a single SVM had only small effects on its accuracy
And Tavares, if nothing's predictive, combining them won't help