Dorothea Floris “Linking structural and functional imaging modalities to characterize face processing in autism”

Name: Dorothea Floris “Linking structural and functional imaging modalities to characterize face processing in autism”
Uploaded: 2023-03-09T21:53:31.2133333Z
Duration: 29 min 20 s

March 09, 2023

Information

ID: 9632
To Cite: DCA Citation Guide

Download Transcript

00:06Good afternoon, everyone.
00:07Welcome to the opera sheet.
00:11And so I'm also very excited to
00:13be back at this great meeting
00:15after the three-year break.
00:17And I'm also excited to present
00:18some of my work here today,
00:19which is on multimodal
00:21analysis and autism within the
00:23context of phase processes.
00:28OK. So autism is a very common
00:31neurodevelopmental condition.
00:32So it occurs in roughly like one person,
00:34one in 44 people,
00:36and it is characterized by social
00:38communicative difficulties and
00:40restricted repetitive behaviors.
00:42And autism is also still exclusively
00:45diagnosed based on behavioral symptoms,
00:48which is pretty much the case for any
00:50psychiatric condition as we know.
00:52But because of this,
00:53our goal is to eventually develop
00:56biomarkers that can help us.
00:58Diagnose the condition more objectively
01:00and also to identify more targeted
01:03support services which help individuals.
01:07And for this reason,
01:08it's of course really important
01:10that we better understand the
01:13neurobiological underpinnings of the
01:15condition and its clinical features.
01:17And in this context,
01:19there has been a lot of neuroimaging
01:22studies conducted already to date
01:25and so many different neuroimaging
01:28modalities have been applied that try
01:31to characterize different biological
01:33aspects of the condition and its
01:37associated features individually.
01:39But then the question is,
01:41can we maybe obtain a more holistic
01:43understanding of the neurobiology
01:45of autism if we study different?
01:48Modalities in combination.
01:52This is what happens on
01:55transatlantic flights. OK.
01:58So luckily it is quite normal that we
02:03acquire multimodal neuroimaging data
02:05and like MRI data from the same subject
02:09when we conduct new imaging analysis.
02:12And the reason why we do so is because
02:15we know that every modality captures
02:18like different unique aspects of the
02:20neural organization of the brain.
02:22So for example,
02:23EG has very high temporal resolution and
02:26most accurately reflects neuronal activity.
02:29And then MRI is you know,
02:31has very high spatial resolution
02:33and provides more detailed insights
02:35into the structural and functional
02:37organization of the brain.
02:39And then depending on our research question.
02:43Um, we study these different aspects
02:45either individually or in combination.
02:48But what we usually do is,
02:49is we conduct unimodal analysis.
02:52So we study each modality individually
02:55and search for like disease or
02:57like phenotype related changes
02:59in each modality separately.
03:05However, we know that like any
03:08single imaging, modality alone can
03:10only provide a limited view into the
03:13neural organization of the brain,
03:15and there should be certain benefits to
03:17combining different imaging modalities.
03:19And so the most obvious benefits
03:21are first of all,
03:22it should be biologically more valid,
03:24of course, because we know that
03:26neural changes occur across different
03:28biological systems at the same time.
03:31Theoretically,
03:31we would also expect a greater signal
03:36to noise ratio and like greater
03:39sensitivity to detect effects.
03:42And it should also give us
03:43probably like a more comprehensive
03:44view of the question we study.
03:54So this figure I borrowed from one of
03:56Minsk Yahoo's papers and it shows you
03:58that there are different approaches
04:00to doing multimodal analysis and
04:01depending on which approach we adopt,
04:03we can increase the amount of the
04:06shared joint variance that we extract
04:08from the different modalities.
04:10So the most simple or like the simplest
04:13way would be to do unimodal analysis
04:15and then visually compare our unimodal.
04:18Results or like do unimodal analysis
04:20and then overlay our results.
04:22So these of course can be very useful,
04:24but they would not account for any
04:26interactions between the modalities.
04:28The next level would be asymmetric
04:31fusion and that's when one modality
04:33is used to restrict another one,
04:36like for example in F MRI seated
04:39EG reconstruction or like when
04:41we Co register images.
04:43And then finally the most powerful
04:45technique would be symmetric fusion,
04:48where we can extract most of
04:50the shared common variants.
04:51So this means that we would merge the data
04:55before the statistical interpretation
04:57stage and model the cross subject
05:00variability in a data-driven way.
05:05And so there are different methods
05:07of course to do multimodal analysis.
05:09So earlier Maria showed us how you can
05:12do structure, function and coupling.
05:14And then what we usually do at
05:16Donders is so-called linked
05:19independent component analysis.
05:20So the toolbox for this was
05:22developed by Alberto Yera at donors,
05:24also Christian Beckman.
05:26And as you can see,
05:28it's pretty much an extension
05:30of single modality ICA,
05:32which you know is like a
05:34multivariate data-driven method
05:36to decompose the MRI data into
05:39statistically independent components.
05:41And different spatial patterns
05:43and associated time courses.
05:44And so here the difference is
05:46it's a Bayesian tensor extension
05:49of single modality ICA,
05:51with the difference that we decompose
05:54the data simultaneously across the.
05:56Different modalities.
06:01Yeah. And as a result,
06:02we would then get independent components
06:04and these are then associated with
06:07spatial maps for each modality,
06:09but with one shared subject course.
06:11And this is the key feature
06:12that the subject course is
06:14shared across the modalities.
06:15And this then you can use to like
06:17relate it to different clinical
06:18features or like behavioral measures.
06:23OK, so few years ago Alberto already
06:26published a paper using this method and
06:28showed how about a powerful tool it is.
06:30So here he applied linked independent
06:33component analysis to all structural
06:35imaging modalities in the HCP data set.
06:37And he found a set of different
06:40independent components and that of which
06:42many related to like the whole range of
06:44behavioral measures that we have in HCP.
06:46And here for example,
06:47you can see one very significant
06:49independent component.
06:51And so these maps are the spatial.
06:53Steps per modality as you can see.
06:56So the interesting thing here was that
06:58all modalities that he fed into the
07:00model contributed to this component
07:02and it was also very significantly
07:04related to a whole range of different
07:06cognitive behavioral measures such
07:08as like cognitive functioning,
07:10mental health, things like that.
07:12So this is just to briefly show you
07:14that it works nicely across structure
07:16structural imaging modalities and
07:18it's a great tool to elucidate brain
07:21behavior relationships across modeling.
07:26But then, as I said,
07:27I'm interested in studying the
07:29neurobiology of autism and so I'm more
07:32interested to apply this method in
07:34a clinical context and characterize
07:36the neuro phenotype of autism cross modally.
07:39And luckily we have a very comprehensive
07:42and deeply phenotype data set available
07:45within the EU aims aims to trials consortium.
07:49So this is the largest consortium
07:52in Europe designed to discover
07:54biomarkers and drugs and autism.
07:56And this is a data set available with
07:59over 700 autistic and non autistic
08:01individuals and at the same time
08:03many different imaging modalities.
08:05So structural MRI DTI is you can
08:08see resting stata from our task
08:10fMRI we also have EG and then like
08:13a whole range of different clinical
08:15and cognitive measures available.
08:18So it's really one of the most
08:21multimodal datasets and largest
08:22ones available in autism.
08:24So it's actually ideally suited to do
08:27multimodal imaging analysis in autism.
08:32OK. So the first study I would like to
08:35touch on and here we asked the question,
08:37what does the multimodal neural
08:39signature of autism look like?
08:41And here we didn't have
08:43any specific hypothesis.
08:44We just looked at like the
08:47global cross modal pattern.
08:49And this study was led by one
08:50of our PhD students at Donners,
08:53Leonard Oblong.
08:55And so here we use this UAMS data set
08:57that I just showed you and integrated
09:00imaging data across three different
09:02modalities which were structural MRI,
09:04resting state F, MRI and DTI.
09:07All analysis and features were
09:10also restricted to specific
09:12cortical and subcortical ROI.
09:14Which had previously been
09:16implicated in autism, such as,
09:18for example, post central gyrus,
09:19amygdala, fusiform gyrus.
09:22And then we did the unimodal
09:24feature extraction which was,
09:25as you can see,
09:26BM for the structural domain,
09:28then connect topic mapping for
09:30resting state fMRI and probabilistic
09:33tractography for DTI and then
09:35went on and applied linked ICA.
09:37To merge these different modalities
09:39and then eventually you would
09:40obtain the subject courses that
09:42are shared across the modalities
09:43and these can then be studied in
09:45association with behavior for example.
09:49OK. So on the bottom you can see the
09:53different independent components.
09:55These are now 22 in total.
09:58The color stands for one
10:00modality and so depending on
10:01the component that we look at,
10:03you can see that there are
10:06more or less multimodal.
10:07Is 22 were the ones that were either
10:10related to some behavioral measures
10:13like adaptive daily living skills,
10:17autism associated features,
10:18or they showed a significant
10:20group difference.
10:25And so the most interesting one among these
10:29independent components was this one here,
10:32because it showed a significant
10:33group difference between autistic
10:34and non autistic individuals.
10:36Or the autistic group as you can see here,
10:37like a lower contribution on this component.
10:41And the next you can.
10:43Look at the spatial profile and the
10:46individual modality contributions
10:47and as you can see we BM contributed
10:49to a very small extent and this
10:52was mostly driven by resting state
10:54fMRI within the fusiform gyrus.
10:58So, taken together in this study,
11:01the most interesting result was not in
11:04a particularly multimodal component,
11:07but it was still interesting to
11:09see this group difference within
11:12the fusiform gyrus and also some
11:15nominally significant associations.
11:17But yeah.
11:20The reason why we were excited to
11:23see this strong implication of the
11:26fusiform gyrus is because it has been
11:29atypically implicated in autism before.
11:32So there are many unimodal studies
11:34that show that autistic individuals,
11:36for example, show hyperactivation
11:37while doing face matching tasks,
11:40or they would show like a delayed and
11:44170 response in response to phases
11:47when doing EEG acquisition or like.
11:50A structurally and they for example,
11:52atypical asymmetry or like increased or
11:55decreased volume depending on the region.
11:58So it has typically been implicated
12:02in autism.
12:03And another important thing is
12:05that it's a key feature associated
12:07with face processing as we know.
12:10And so you can see here that it's kind
12:12of like a mosaic with like different
12:15differently functionally specialized regions.
12:17So there are these like
12:19circumscribes patches like for.
12:21That are like responsible for places,
12:23processing shapes,
12:24words and faces,
12:26but it's like the key region that's
12:29active when you process phases.
12:31And what's also important here is that
12:34atypical phase processing has been shown to
12:37be one of the most commonly cited social
12:40difficulties in autistic individuals,
12:42who they Orient much less to faces.
12:45But they also have difficulties understanding
12:48or like recognizing facial emotions.
12:51OK.
12:51So this is just a brief overview
12:54or context of the fusiform gyrus.
12:56So the next study makes more sense
12:58because based on Leonard's results.
13:00And based on these unimodal prior
13:03results or literature in autism
13:05and I decided to apply this linked
13:08ICA method to the fusiform gyrus
13:10specifically and asked what does
13:12the multimodal neural signature
13:14of phase processing look like in
13:16the fusiform gyrus and autism?
13:20So here I also merged them
13:23different imaging modalities,
13:24so also structure and resting
13:26state fMRI as Leonard did,
13:28but it also included additionally 2
13:30functional modalities that were specifically
13:32associated with phase processing,
13:34so EG and task fMRI.
13:38So for EG there was ERP data available
13:43for those recorded when subjects looked
13:46like upright or inverted phases and
13:48for task fMRI this was acquired when
13:52subject performed the Hariri face
13:55matching paradigm in the scanner.
13:57And I told you that this UM sample
14:00is actually quite big that but then
14:02when we take the intersecting sample
14:04across all the different modalities,
14:06that's available for all the
14:08individuals and also with good quality.
14:10This actually boils down to a
14:12sample of around 200 individuals.
14:16OK, so we can look at the
14:20different methodological steps.
14:21First of all, I restricted all
14:23analysis to the left and the right
14:25fusiform gyrus separately because
14:27we know face processing is the right
14:29lateralized cognitive function and
14:31it actually makes sense to study the
14:35hemispheric contributions individually.
14:37Then for the structural domain,
14:39I am extracted Gray matter
14:42volumes based on VM.
14:43And for task F MRI we used the
14:47contrast maps for the phases
14:50greater than shapes condition from
14:53this Hariri phase matching task.
14:56And for resting state F MRI I computed
15:00connectivity gradients or connect topic maps.
15:04And this is actually the same approach that
15:07Michael introduced before in the hippocampus.
15:10So you correlate each voxel within
15:12the fusiform gyrus with the voxels
15:14and the rest of the cortex.
15:16Then you can compute the similarity
15:20matrix and derive Laplacian eigen maps.
15:24And this then reflects the connectivity
15:26gradient within the fusiform gyrus
15:28and it it makes actually sense to use
15:31a spatial model like this because
15:33of this fine grained functional
15:35heterogeneity within the fusiform gyrus.
15:39OK, and then for EEG,
15:42we did source reconstruction
15:44with interviews of from gyrus,
15:46so we reconstructed them the time
15:48series from different locations
15:49within the fusiform gyrus.
15:56Then as a next step,
15:57I applied normative modeling to
15:59each imaging modality separately.
16:01And so we have the true experts on
16:03normative modeling from the donors
16:05in the audience, Charlotte and Sage.
16:07And So what I did here was I computed
16:10the mapping between the brain
16:13features and age, sex and site.
16:15And with this we derived the values
16:18which quantify how much each modality and
16:22differs or like deviates from a normative.
16:25Pattern at the box lower at the
16:28time series level, time Point left.
16:32And then next these Z values or deviations
16:36were fed into the linked ICA model.
16:40So this is where you merge the different
16:42modalities using linked ICA and as I said,
16:45as a result we get how much each
16:47modality contributes and also how
16:48much each subject contributes.
16:50And this is this cross subject variability
16:52that we can then also use to study it
16:56and association with behavior or like
16:58some clinical features of interest.
17:03OK, so here you can see the
17:06resulting independent components.
17:07These are the multimodal ones.
17:09So linked ICA would also
17:12theoretically give you unimodal ones.
17:13But here I'm interested in
17:15the multimodal components,
17:16which are the ones where one modality
17:19does not contribute more than 80%.
17:21As you can see,
17:22they're also divided into right and
17:24left hemispheres separately and sorted
17:26by how multimodal they actually are.
17:29We can say that the right in
17:31the left hemisphere had like
17:33roughly equal contributions here.
17:35And overall, EG had the largest contribution,
17:38whereas VBM had the smallest contributions.
17:41Next weekend and also check what the
17:45functional meaning or implications
17:47of these multimodal components are.
17:50So, for example,
17:52we can check how these multimodal
17:55components relate together.
17:58So not individually,
17:59but like in the spirit of
18:01multimodal analysis,
18:02how they relate together to different
18:04cognitive constructs that are
18:06relevant in the context of social
18:09functioning and face processing.
18:10And so a way to do this is by a
18:13canonical correlation analysis, which.
18:16Basically identifies a multivariate
18:19association between brain related
18:22features and like a set of social
18:25communicative features that we choose.
18:28And as you can see here,
18:29this resulted in a significant
18:32association between between
18:33the two sets of variables.
18:36You can also see that they
18:38differently load onto.
18:41This result so some independent
18:43components loaded more than others.
18:46The same applied to the social
18:48cognitive features.
18:49And what was also interesting was
18:51when we reran this with like other
18:54measures related to restricted
18:56repetitive behaviors and like sensory
18:58processing like non social features
19:01this was no longer significant.
19:02So it points to some specificity
19:05in the social communicative
19:07domain and phase process.
19:11OK. And then we can check our any
19:13component significantly different between
19:15autistic and non autistic individuals.
19:17And there was one component that
19:19significantly differed where autistic
19:21individuals had significantly lower
19:23contributions than non autistic individuals.
19:26This was a component that was driven
19:29by task fMRI, resting state fMRI,
19:31EEG and also to a small extent VBM.
19:34So actually all modalities contributed
19:36that we fed into the model just
19:39to like varying degrees. Umm.
19:44Here also the right hemisphere had a larger
19:47contribution which is interesting in the
19:49context of like the lateralized phase
19:51processing effects that we would expect.
19:53And you can also see the different
19:56spatial maps for each modality and
19:58the different regions within the
20:01fusiform gyrus that load positively
20:03or negatively onto this component.
20:05And then so here in this case where we
20:08see the significant group difference.
20:12We can then say that in the
20:14positive yellowish regions,
20:15autistic individuals loaded more,
20:18had lower loadings,
20:20whereas in the bluish regions
20:21higher loadings onto this component.
20:23And then depending on the modality we
20:26could interpret this accordingly that
20:28they had like larger or like smaller
20:30volume or higher or lower activation.
20:33Hmm.
20:35And then we can further characterize
20:38this component by looking at the
20:41different modality contributions.
20:43For example, if you look at the temporal
20:45profile of EG and the different time
20:48points that most strongly load onto.
20:51This component we can see that this
20:53was the case at like 100 and 72180 and
20:58450 milliseconds.
20:59And this was quite interesting
21:01because we know that for example the
21:03face sensitive ERP occurs at N 170.
21:06Which is associated with phase processing
21:09and expertise to social stimuli and
21:14between 280 and 500 milliseconds for
21:17example we see the P300 and it's
21:20sub components P3AP3B and these also
21:22have been associated with learning,
21:24novelty detection things like that.
21:27So yeah I'm not big expert but it
21:29was interesting to see that this
21:32inter subject variability and
21:33like group differences occurred
21:35at these time points that.
21:36Are actually meaningful.
21:39And then for the the remaining modalities,
21:43we can further look into
21:45the spatial profiling,
21:46characterize these for example
21:47overlay them with an Atlas
21:48like the Harvard Oxford Atlas.
21:50And then we see, OK,
21:51most of the group differences occurred
21:53mostly like in more posterior regions
21:56or like posterior occipital regions.
21:59You could also overlay it with the.
22:01Functional Atlas.
22:02So there's this visual functional Atlas.
22:06Which was created,
22:07so they ran different functional
22:09localizer task and so it does.
22:11Atlas was then created and characterizes
22:14this functional heterogeneity within
22:16the fusiform gyrus and describes
22:18these different categories,
22:19specific category specific patches and
22:21we can see that there is some overlap.
22:24So for example they autistic individuals
22:27show decreased volume in like more
22:29in the more retinal topic areas of
22:31the fusiform gyrus also in the right.
22:34To the form face area,
22:36but the initial increased values for
22:39the resting state modality, for example.
22:42So this probably shows us that it's
22:45more like the posterior regions,
22:47but like that are involved more in
22:49early visual processing are involved,
22:51but also the fusiform face area,
22:53which is of course very.
22:54Interesting in this context.
22:58OK.
22:58So in summary,
23:00we can say that we successfully
23:02merged data across different new
23:05imaging modalities to characterize a
23:08multimodal neural phenotype of autism.
23:11Multimodal aspects related to the
23:13fusiform gyrus and phase processing
23:16are related to different behavioral
23:18phenotypes and can explain variants
23:21and social functioning and phrase
23:24processing and autism.
23:25We can also say that multimodal
23:27approaches are useful in general
23:29because theoretically they should bring
23:32us closer to biological validity.
23:34As we know that they said the changes
23:37occur across different biological systems,
23:40it should also increase our signal
23:42to noise ratio,
23:43so we might be more sensitive to
23:45detect effects as we saw earlier
23:49in josephina's talk.
23:50I think she also showed that the the
23:53multimodal feature classification.
23:54Outperformed the unimodal one if I
23:58remember correctly. So this is one example.
24:00But then here we could also compare the
24:03unimodal with the multimodal analysis to,
24:06it's like Randy said, to quantify,
24:08quantify actually the added value
24:12of doing these multimodal analysis.
24:15Because theoretically, you know,
24:17unimodal ones are easier, faster,
24:20like might be more economic.
24:23But still,
24:24usually we collect all these
24:25imaging modalities,
24:26so if we have the possibility,
24:28we should aim for doing this,
24:29especially in our big data.
24:32Error and eventually this should
24:34help us to better characterize
24:37the neurobiology of autism or
24:40any other neurodevelopmental
24:42condition. And yeah, so. Go merch.
24:54And they would like to thank, of course,
24:55everyone who contributed to this work
24:57from University of Zurich and donors and
24:59the UIMS consortium and thank you guys.
25:03Last question.
25:07So this is really cool.
25:09This is important.
25:10I wonder if this approach allows you to.
25:14Gaining insights into like mechanistically
25:17being put allergies or if you see it more
25:20as a as a kind of these are combined.
25:25Get better. No.
25:29Yeah. So the question is
25:31what it's most useful for,
25:32more like for a mechanistic insights or
25:34more for increasing our predictive power.
25:36I'd say both. So definitely you
25:40know when you do unimodal analysis.
25:42It can be useful of course too,
25:43and you get like more the salient
25:46features of the one imaging modality.
25:49But you know, it could be completely
25:51different systems involved when you look at,
25:53you know, the different ones that you
25:55would not detect when you look at one and.
25:58More like you know that they
26:01interact in combination when it
26:03comes to the different disorders.
26:04So yeah, when we don't also combine it with,
26:06for example, some.
26:08Genetic analysis,
26:09you could do like gene expression decoding,
26:11you know in these regions that
26:13are cross modally implicated,
26:14we could get some more mechanistic
26:17insights and then as we saw earlier it
26:19can also increase our predictive power.
26:23So yeah.
26:27Very interesting analysis,
26:28very interesting results.
26:30Can you help me get the
26:31head around one thing?
26:34In the linked ICA the component that's only.
26:39Hmm. Then still influenced
26:43by the other modality.
26:46Yeah, that's a good question.
26:47So you could compare compare it
26:49with like unimodal ICA to see if
26:50you get them the same pattern.
26:52But that's the way that linked ICA works.
26:55It has its uses like this
26:58Bayesian model order selection,
27:00so it has like these are ARD prior, so like.
27:04Automatic relevance detection.
27:06So it actually gives each modality away.
27:09Weighting, weighting,
27:10and then can eliminate modalities that are
27:13not informative for that one component.
27:16So it can also identify,
27:18you know, structured, structured,
27:22single modality signals.
27:24If there are also the same ones that
27:26would come up if you do single modality,
27:29I say. We can check would that be?
27:37You separate out everything from signal
27:39that also relates to the resting state
27:43and this is like what's left over there.
27:45So this one would be a unimodal
27:47component and I mean then it's not
27:49driven by the other modalities.
27:51So that's then it can also be for example
27:53noise associated with this one component.
27:55You know you can check is it related
27:58to head motion or other confounds.
28:00So it can get these single modality
28:02structured signal or noise components
28:04but then unrelated to the other.
28:06Modalities. Yeah.
28:14For study, you showed the different
28:16aspects of the present from giant
28:18technological data related to this one
28:21component that wasn't system mean.
28:24That's what I was trying to figure
28:27out with these last slides.
28:29It's complicated, yeah,
28:30but it's like the regions that are
28:32cross modally implicated and then you
28:34can like characterize these further.
28:35You know, do they more overlap with like
28:38the face from face area or you know,
28:40do we see more like object place
28:42related regions and like which could
28:44indicate that autistic individuals
28:46have like an atypical strategy even
28:48processing phases or is it more the
28:50left versus the right, you know,
28:51we would expect more right to the.
28:54Involves more to the opposite hemisphere,
28:56so you can make sense of these
28:58different patches,
28:59but like individually by modality then.
29:05Or only through? Thank you. Yes well
29:09here it's this connect topic maps only.
29:13Yeah.
29:17OK.
29:20OK.