NBE-E4530 Special Course in Human Neuroscience Human brain connectivity (Spring 2017) Enrico Glerean twitter: @eglerean
Connectivity with fmri: from preprocessing to networks. Tools for resting state and task connectivity. Enrico Glerean twitter: @eglerean
Connectivity with fmri: from preprocessing to networks. Tools for resting state and task connectivity. Types of fmri experiments and approaches to connectivity Preprocessing optimised for connectivity Sources of artifacts
The brain at rest! LETʼS SLEEP FOR A BIT! Enrico Glerean! www.glerean.com @eglerean!!
The activity of the brain at rest is ideal for estimating the connectome" By looking at regions that change together in time we can estimate their connectivity! Raichle, M. E. (2010). Two views of brain function. Trends in Cognitive Sciences, 14(4)" Enrico Glerean! www.glerean.com @eglerean!!
Experiment design in neuroscience
I cannot stress this enough: with a shitty experiment, you get shitty data Good data collection starts with great experiment design The field is still struggling to convince everyone that experiment design is very important http://biorxiv.org/content/biorxiv/early/ 2017/03/23/119594.full.pdf http://biorxiv.org/content/early/ 2016/09/25/077131
I cannot stress this enough: with a shitty experiment, you get shitty data There should be a whole course dedicated to experiment design But we are busy people, so it is not wrong to copy what others did before us. I.e. do not re-invent broken wheels, there is a reason why there are just a handful of designs
Reminder fmri
Functional magnetic resonance imaging (fmri)" We measure multiple time series at once" We can consider them independently (e.g. GLM) or we can look at mutual relationships! Blood Oxygen Level signal! Enrico Glerean! www.glerean.com @eglerean!!
What to do during a brain scan?
What would you ask your subject to do in a brain scanner?! (BE CREATIVE!)!! presemo.aalto.fi/hbc2017/!
What should a subject do in the scanner with fmri? Presence of a task: Task should be matched with the signal you measure. With BOLD you must ask yourself What we can do and what we cannot do with fmri http://www.nature.com/nature/journal/v453/n7197/ abs/nature06976.html Lack of task: spontaneous brain activity. More related to anatomy.
What is the subject doing? 1. The subject is doing a task Multiple options 2. The subject is resting A. with eyes closed I. awake II. asleep B. with eyes opened
Break Back here at
What is the subject doing? 1. The subject is doing a task 2. The subject is resting A. with eyes closed I. awake II. asleep B. with eyes opened
Resting state 1. The subject is doing a task 2. The subject is resting A. with eyes closed I. awake II. asleep B. with eyes opened
Resting state PROs The most simple paradigm, works with any healthy or clinical subject as long as they can stay still in the scanner, strong overlap with structural connectivity. Can become a biomarker. CONs There is no such thing as free lunch, we don t know what the subject is doing (asleep? Ruminating?), data driven -> watch out for false positives due to artifacts (motion, breathing, heart, )
Resting state, what art thou? Functional connectivity between surgically disconnected regions? http://blogs.discovermagazine.com/neuroskeptic/2017/04/21/functionalconnectivity-discon-fmri Resting state literature is sleep literature Analyzing 1,147 resting-state functional magnetic resonance data sets, we revealed a reliable loss of wakefulness in a third of subjects within 3 min http://www.cell.com/neuron/abstract/s0896-6273(14)00250-5
Building a functional network" " At each node we measure a time series We compute their similarity! b 1 (t)" b 2 (t)" Enrico Glerean! www.glerean.com @eglerean!!
Building a functional network" Similarity value used as weight of the edge between the two nodes. Repeat for each pair of nodes." r 12! b 1 (t)" r 12! e.g. Pearsonʼs correlation:" r 12 = corr(b 1 (t),b 2 (t))" b 2 (t)" Enrico Glerean! www.glerean.com @eglerean!!
Resources for resting state Conn toolbox (Matlab) https://www.nitrc.org/projects/conn/ REST toolbox (Matlab) http://restfmri.net/forum/index.php Python s nilearn http://nilearn.github.io/connectivity/ index.html
What if same noise is in both time series?" If the subject move suddenly for example, we have a spike in both time series -> strong correlation" r 12! b 1 (t)"??????????! e.g. Pearsonʼs correlation:" r 12! r 12 = corr(b 1 (t),b 2 (t))" b 2 (t)" Enrico Glerean! www.glerean.com @eglerean!!
Task 1. The subject is doing a task 1. Task structure 1. In Blocks 2. As Events separated in time 3. As a stream of events (naturalistic) 2. Passive vs Active 1. Pressing a button, etc 2. Just watching and mentalizing
Book: Sarty Compu'ng Brain Ac'vity Maps from fmri Time- Series Images Course: h@ps://www.coursera.org/learn/func'onal- mri Enrico Glerean! www.glerean.com @eglerean!!
How to analyze task connectivity given task structure The more structured the task, the less you can use the time series (and viceversa) With block and with (not too fast) event related design we use the general linear model GLM to abstract from the time series into activations
How to analyze task connectivity given task structure With 20s blocks, the best is PPI* Y = (A/- NoA/) β 1 + V1 β 2 + (A/- NoA/) * V1 β 3 + e Modeling signal Y, given task, given another signal (V1), given an interaction between task and signal Source: h/p://www.fil.ion.ucl.ac.uk/mfd_archive/2011/page1/mfd2011_conneclvity_ppi_sem.pptx *PPI = psychophysiological interaclon
How to analyze task connectivity given task structure Resources for PPI SPM (matlab) FSL (stand alone) gppi (generalized PPI, https://www.nitrc.org/projects/gppi) *PPI = psychophysiological interaction Source: h/p://www.fil.ion.ucl.ac.uk/mfd_archive/2011/page1/mfd2011_conneclvity_ppi_sem.pptx
How to analyze task connectivity given task structure Event related, the best is beta series For every event we compute a beta weight in the GLM sense We replace BOLD time series with beta time series We correlate beta time series between regions
How to analyze task connectivity given task structure
How to analyze task connectivity given task structure Resources for beta series https://www.ncbi.nlm.nih.gov/pmc/articles/ PMC4019671/ BASCO toolbox: https://www.nitrc.org/projects/basco/ Mini function I made: https://version.aalto.fi/gitlab/bml/bramila/ blob/master/bramila_betaseries.m Source: h/p://www.fil.ion.ucl.ac.uk/mfd_archive/2011/page1/mfd2011_conneclvity_ppi_sem.pptx
Task 1. The subject is doing a task 1. Task structure 1. In Blocks 2. As Events separated in time 3. As a stream of events (naturalistic) 2. Passive vs Active 1. Pressing a button, etc 2. Just watching and mentalizing
Correla'on approaches Let s consider two 'me series for two voxels b 1 (t) b 2 (t) Enrico Glerean - Brain & Mind Laboratory Aalto University School of Science (Finland)
Correla'on approaches Let s take all 'me points b 1 (t) b 2 (t) Enrico Glerean - Brain & Mind Laboratory Aalto University School of Science (Finland)
Functional connectivityin time Sliding window correlation b 1 (t) w n b 2 (t)
Functional connectivity in time Sliding window correlation for functional connectivity produces link time-series b 1 (t) r 12 (n) w n b 2 (t) e.g. r 12 (n) = corr(b 1 (w n ),b 2 (w n ))
Problems with sliding window connectivity Field is still arguing what Dynamic Functional Connectivity means Size of window depends on the temporal frequencies of the signal http://www.sciencedirect.com/science/article/pii/s1053811914007496 https://www.ncbi.nlm.nih.gov/pmc/articles/pmc4758830/
Functional connectivity in time: other approaches Wavelet decomposition https://www.ncbi.nlm.nih.gov/pmc/ articles/pmc2827259/ Multiplication of derivates http://www.sciencedirect.com/ science/article/pii/ S1053811915006849 Phase synchronisation (Glerean et al 2012) https://www.ncbi.nlm.nih.gov/ pubmed/22559794
fmri preprocessing optimised for FC
Itʼs 2017! AND STILL WEʼRE TALKING ABOUT FMRI PREPROCESSING?!! Enrico Glerean! www.glerean.com @eglerean!!
Know thyself! OR ACTUALLY KNOW YOUR DATA!! Enrico Glerean! www.glerean.com @eglerean!!
How awful is the BOLD signal?" BOLD response on same subject has different lags in different days (Aguirre, 1998, Neuroimage)" Caffeine changes BOLD and resting state networks (multiple studies)" Eating salad (= NO3- nitrate intake), changes lag and amplitude of BOLD response (Aamand et al 2013, Neuroimage) hint: TELL SUBJECTS TO EAT SALAD!! Scanner parameters matter: different activations to same paradigm and same subject (Renvall et al 2014 Scientific Reports)! " Enrico Glerean! www.glerean.com @eglerean!!
How awful is the BOLD signal?" Even within same subject and same moment, BOLD response does not have the same lag/amplitude across all voxels! fmri has bad reputation because of too many study overlooking confounds and over-interpretation of results (often no replication with other imaging modalities e.g. M/EEG) [see http://neurochambers.blogspot.fi/2014/01/tough-love-for-fmri-questionsand.html]" I am glad I asked this blogger to collect all the sources of confounds with BOLD https://thewinnower.com/papers/concomitant-physiologic-changes-as-potentialconfounds-for-bold-based-fmri-a-checklist " Enrico Glerean! www.glerean.com @eglerean!!
Despite all this! WE CAN STILL CONVINCE OTHER SCIENTISTS THAT WE CAN DO RELIABLE FMRI STUDIES BY CONTROLLING FOR CONFOUNDS!! Enrico Glerean! www.glerean.com @eglerean!!
fmri preprocessing in 1 minute
Preprocessing in 1 minute Raw data as collection of slices interleaved grouped in volumes Estimate motion in xyz space (3 translation and 3 rotations) and correct for it Coregister individual brain to standard brain space Sometimes, spatially smooth the data (reduces inter-individual differences) Sometimes, temporally smooth the data (reduces some known artefacts)
Noise!
Noise in the BOLD signal spectrum" Physiological noise but also scanner drift! Having the same noise in two time series will cause them to be similar because of the noise! Ideally you should de-noise the data, often this is not possible and data is band-pass filtered (0.01 0.08 Hz)! Source: Glerean et al 2012, Brain Connectivity doi:10.1089/brain.2011.0068" Enrico Glerean! www.glerean.com @eglerean!!
Motion!
How does motion affect MRI?" Anatomical images (MPRAGE)" - Long acquisition time, you see that the slices do not align at some point if the subject moves in the middle" Diffusion imaging (DTI & DSI)" - Even longer acquisition time, motion will corrupt one (or more) of the directions you sample. See recent PNAS paper on lower FA in ASD explained by head motion " Functional images (EPI)" - Faster acquisition time -> less part of the data is affected" - Interleaving slices mitigates the problem a bit" - Spatial and temporal smoothing can cure mild artefacts " - Art-repair (interpolating corrupted volumes) could be used" - however any bi- or multi- variate method will be strongly affected! Enrico Glerean! www.glerean.com @eglerean!!
fmri! QUALITY CONTROL OF HEAD MOTION! Enrico Glerean! www.glerean.com @eglerean!!
Quantify amount of motion" Enrico Glerean! www.glerean.com @eglerean!! "Power et al 2011, Neuroimage"
Quantify amount of motion" "Power et al 2011, Neuroimage" Given the rigid body parameters" Estimate the Framewise Displacement at each time point i! Given a subset of voxels or ROIs" Estimate the DVARS (Derivative of VARS = variance over voxels) [Warn: handle with care with non rest data] " Enrico Glerean! www.glerean.com @eglerean!!
Motion and fmri! WHAT USUALLY HAPPENS AND HOW IT LOOKS LIKE IN THE SIGNALS! Enrico Glerean! www.glerean.com @eglerean!!
"Power et al 2014, Neuroimage"
Motion and fmri! WHAT YOU DIDNʼT EXPECT TO HAPPEN AND HOW IT LOOKS LIKE IN THE SIGNALS! Enrico Glerean! www.glerean.com @eglerean!!
fmri FC! FUNCTIONAL CONNECTIVITY! Enrico Glerean! www.glerean.com @eglerean!!
Functional connectivity" Similarity value used as weight of the edge between the two nodes" r 12! b 1 (t)" r 12! e.g. Pearsonʼs correlation:" r 12 = corr(b 1 (t),b 2 (t))" b 2 (t)" Enrico Glerean! www.glerean.com @eglerean!!
Motion and FC! HEAD MOTION GENERATES SPURIOUS BUT SYSTEMATIC FC RESULTS! Enrico Glerean! www.glerean.com @eglerean!!
Mitigate impact of motion on FC: Scrubbing" Once you identified the bad time points, you do scrubbing: i.e. you avoid considering them in your pairwise correlations (discard subjects that had less than 5 minutes left) " "Power et al 2011, Neuroimage" Enrico Glerean! www.glerean.com @eglerean!!
Effects of motion on FC " In most cases scrubbing recovered long distance connections" In few cases it mitigated false positives! Short distance connections are basically untouched" "Power et al 2011, Neuroimage" Enrico Glerean! www.glerean.com @eglerean!!
BraMiLa tools! HOW TO REDUCE SPURIOUS EFFECTS DUE TO MOTION IN YOUR FMRI DATA! Enrico Glerean! www.glerean.com @eglerean!!
BraMiLa: implementation of Power et al 2014, Neuroimage" Enrico Glerean! www.glerean.com @eglerean!!
"Power et al 2014, Neuroimage"
"Power et al 2014, Neuroimage"
BraMiLa implementation" /scratch/braindata/shared/toolboxes/bramila/ We have a GIT repository if you want to help!
Open issues! IS THIS PIPELINE ENOUGH? (HINT: NO)! WHAT ABOUT GLOBAL SIGNAL REGRESSION?! Enrico Glerean! www.glerean.com @eglerean!!
Conclusions by Power et al. 2014" " Enrico Glerean! www.glerean.com @eglerean!!
What about task related fmri?" BOLD GS correlates with task (*)" BOLD from WM correlates with task (*)" BOLD from CSF signal correlates with task (*)" (*) Van Dijk et al., 2010, Journal of Neurophysiology " Head motion can correlate with task and - by the way - with phenotypic traits (Men move more than women!, Van Dijk, 2011, Neuroimage) " Heart beat and breathing correlate with task" itʼs a hell of a mess what can we do about it?" Enrico Glerean! www.glerean.com @eglerean!!
Task related fmri: data health checklist" Check that motion parameters, heart rate and breathing rate are not intersubject correlated! Check that motion parameters, heart rate and breathing rate are not task correlated" Check that motion parameters (e.g. mean FD) are not correlated with phenotypic (age, sex) or behavioural scores" When doing GS/WM/CSF/HM/HR/BR regression, consider other subjectsʼ signal so that you can identify individual principal components of GS/WM/CSF/HM/ HR/BR" Always use average HM parameters in your group statistics (see Yan et al, 2013, Neuroimage)" 1970 Monty Python!
Final conclusion 1! THE BEST MOTION REMOVAL TOOL IS TO TELL MILLIONS OF TIMES TO THE SUBJECTS TO NOT MOVE IN THE SCANNER!! (DATA IS EXPENSIVE, WE SHOULD CONSIDER PRACTISE SESSIONS FOR NEW SUBJECTS)! Enrico Glerean! www.glerean.com @eglerean!!
Final conclusion 2! INCLUDE ALL CONFOUNDS IN YOUR ANALYSIS (AGE, SEX, ). SHOULD WE START ASKING THE SUBJECTS ABOUT COFFEE, SALAD OR OTHER INTAKE?! Enrico Glerean! www.glerean.com @eglerean!!