Player is loading...
Brain MRI Analysis and Machine Learning for Diagnosis of Neurodegeneration
Dr. Ahmed Abdulkadir, University of Pennsylvania
Thursday, Feb. 6, 2020 · 10:04 a.m. · 48m 54s
Cognitive and neuro-biological profiles of elderly individuals are very heterogeneous. Healthy aging, psychiatric disorders, genetic variability, and neuro-degenerative diseases all contribute to this heterogeneity. Classification systems with categories and sub-categories of neuro-degenerative disorders were traditionally used to identify differences across groups. However, a large portion of the observed variability in clinical manifestation on the individual level remains unexplained. Uncertain diagnostic labels, the presence of possibly multiple interacting conditions, and systematic biases due to technical variations in the acquisition hardware further call for innovative solutions. During his PhD, Ahmed Abdulkadir focused on improving MRI image classification and segmentation using deep (end-to-end) and shallow machine learning. The outcomes from the thesis and the literature corroborated that disruptive innovations in the field will be facilitated by the ability to make predictions with confidence bounds in heterogeneous populations with incomplete data. Methodological grounds and preliminary results of ongoing studies as well as a quick preview on the recently accepted SNF Postdoc.Mobility proposal titled MRI-based characterization of subgroups in aging and dementia will be presented.
Embed
Transcriptions
Note: this content has been automatically generated.
00:00:01
for this a nicer introduction i'm i'm very excited to to
00:00:06
um to be here and i was already told me did about
00:00:09
about the history of the of this institute about it could about
00:00:12
how we'd grew and about the outstanding research that is done here
00:00:17
so my background i started already said he's um more in psychiatry in more um more specifically
00:00:24
i'm kind of apply machine learning in the field of psychiatry and i'm currently kind of
00:00:31
almost in between jobs so essentially um since this week i am a post
00:00:36
doc mobility fellow um pennsylvania in the group off that has a lot to cause
00:00:43
university of pennsylvania and talk to i'm going to send today is kind of should give you a bit of overview
00:00:51
um pollen what i did to during my p. h. d. c.'s is then
00:00:56
what id it kind of the the last year more less or when it's in the first post talk a
00:01:03
face and then also little bit of an outlook into uh what i'm planning to do and they're kind of to
00:01:10
the sauce behind what i'm want to do um in in the two years off postal mobility
00:01:18
so the the first part to fiji project these
00:01:22
divided in in in three main parts which is one is
00:01:24
image classification so this is a new imaging uh classification then
00:01:29
uh the next part is about assisted assisting radiologists
00:01:33
in uh in assessing phrases atrophy in in the brain
00:01:39
in the last uh is about people and in the second
00:01:43
part i will talk about three applications of off the machine learning
00:01:49
and then insert part about the more more rested that outlook
00:01:59
so first couple of words why we're interested in m. r. i. so ever i on like
00:02:05
computer more fit for instance is non invasive that means there is no harm done
00:02:12
by by measuring orbit imaging um the brain so there's there's no um radiation for instance
00:02:18
can also m. rise very versatile so the whole principle isn't based essentially on the principle
00:02:25
of magnetic induction the uh which is a depicted here which is um essentially in the strong
00:02:32
magnetic field off to m. ha or a device the it's kind
00:02:36
of the the figuratively bits not physically not exact like especially with leslie
00:02:40
uh there's a minus taxation vector data lines to this field and then these excited by um
00:02:47
by high frequency field we're right at the frequency and
00:02:50
then this this vector starts to process and then and meet
00:02:55
um if it meets essentially electromagnetic radiation and the
00:02:59
signal when we now know removed a spatial coding
00:03:04
uh is essentially depending on a couple of property so it's depending
00:03:08
on the proton density also tissue uh it's depending on a two
00:03:14
kind of tissue properties that can be kind of simplified yeah into um time
00:03:19
constance constance of this uh exponential decay as a t. one and twenty two
00:03:25
and then the contrast between tissues is essentially obtained by uh the sequence parameters uh
00:03:32
so to t. i. inversion time um chemical time and we cover time so the
00:03:39
just a basic idea here is that by manipulating these tests we uh the kind of
00:03:45
six parameters we can get it from the same brain to
00:03:48
images that look very different in terms of a tissue contrast
00:03:57
and then the big part off uh my my uh digits easiest
00:04:03
if all the rounds actually there's very classical machine learning a
00:04:07
part time where there's some for all data in the top
00:04:11
rope so this can be functional and varieties can be structured
00:04:14
i'm right here then there's some preprocessing done like core just ration
00:04:22
then there is some sort of of feature extraction step that is
00:04:25
based on prior knowledge so for instance what you see here is that
00:04:31
uh so there's like um for so from this grey scale image we obtain a so called to should probably too much that are
00:04:37
based on one hand phone on the intensity of the image but on the other hand
00:04:43
also on the spatial location or or and prior topless that in that case
00:04:47
and then we know that the amount of grey matter in different regions is indicative for for instance a certain pathologies
00:04:55
and then when we have these features and then we do pretty simple pattern recognition uh for
00:05:01
instance uh support vector machines here i show a kind of an example of of such a such
00:05:12
a pipeline to to kind of do automated diagnosis or even specification um with classical machine or not
00:05:21
but we start with a with a native but one image anyhow these um and political template that
00:05:28
guide essentially these uh issue classification um when
00:05:33
we apply princes these pieces of the uh the
00:05:37
light blue collar here uh indicate the grey matter uh and any desperation organisations such that every brain
00:05:44
of every subject all anatomical regions um our align because in that case we can uh when we
00:05:53
smooth image with that we can just do it direct i'm expecting buddies and you just take the
00:05:58
the pixels or the boxes of of the images um as vectors and a half then that
00:06:04
way i get a feature vector with matching uh features across subject um and then we can do
00:06:10
this uh in different uh spatial spatial scales so
00:06:14
because the apology can be either affecting only small
00:06:19
small regional it's a box of or also um entire region and uh with this kind of either
00:06:25
on the box level or on on um and comical region regions we can kind of have
00:06:31
these multiple special skates and then this wouldn't be give 'em 'cause can be combined like in in
00:06:37
some way to speech can become quite made up then from two that was structural m. right from
00:06:45
functional m. writer also some uh the tape classical
00:06:49
paradigms that are used to extract certain and features
00:06:52
and one of the most widely used is uh the like a normal a unit on login generally your
00:06:59
model that essentially um when you observe with the signal
00:07:03
why we assume that the signal is is uh based
00:07:10
on some some model so essentially what we see here is that during this phase for instance there was no
00:07:17
image shown then here some image was shown so during this this phase where the where the model is high
00:07:24
then no image shown and then there's an image shown here and then there's a no image um and the part i'm
00:07:31
here is that there is some activity or blood flow uh during these times here and then the signal that we observe in
00:07:38
different regions of the red and green region would be there
00:07:41
and like essentially scaled versions of this week's obviously um some noise
00:07:47
so when you don't estimate the parameters of of this uh
00:07:50
of the scale here for instance we can kind of a contrast
00:07:56
different activities so for instance uh which regions are active when we see a building versus when
00:08:03
we see a person so then d. kind of one of the first what's that i did
00:08:12
which was not necessarily clinically relevant but was was very important in a in a strategic sense
00:08:18
so at that time we you work kind of evaluating whether to put our efforts in structural imaging
00:08:25
or rather in functional imaging for the diagnosis
00:08:30
of preclinical your degeneration now the problem with preclinical
00:08:34
preclinical means that there are no clinical sites so that means in in uh in in subject
00:08:40
of having no clinical signs we would we can only see if they develop signs your flight
00:08:47
and to to kind of as a model to for print preclinical diagnosis we use the population of
00:08:53
people with um huntington disease is it genetic disease that um based on the kind of genetic
00:09:01
characteristics determines the very very act or quite
00:09:07
accurately when persons will start to have symptoms so
00:09:12
that means we can although this subject did not have any kind of a typical symptoms we
00:09:18
we could we knew from the genetic profile when they would develop such symptoms um and and we
00:09:22
use this as a kind of as as a as a more so and and the basic
00:09:27
was also what you see here we compare uh so you see like three example format controls
00:09:33
and then from the symptomatic a genetic mutation carriers uh on one hand i'm on the a.
00:09:39
u. c. d. at this kind of structural features um and here you see a like functional features
00:09:46
specifically in the region of interest that was suspected to be to be a
00:09:51
relevant for this uh task uh and you for another uh eh from right
00:09:57
um our time and the the result that we essentially obtained is that uh
00:10:05
the functional imaging never consistently outperform the structural imaging and that kind of lead
00:10:13
us to the to the strategic decision to for the focus on on a
00:10:19
structural um camera so the uh the next step uh in in the fees
00:10:28
is uh was that there are some some corvallis like age and sex that
00:10:34
have a definite impact on the buyer marketed measure let's say great but the volume
00:10:38
in so we can and then we just hypothesise that we can try to
00:10:43
just like we move a remove these these uh these effects uh and then
00:10:50
hopefully improve the the classification and then we uh we moved the effects linearly
00:10:55
and also non linearly and there's also an showing now is kind of with these
00:11:01
um with this in mind and the for the nonlinear correction we also wanted
00:11:10
to have a kind of not up not have a strong prior happens is
00:11:16
about how for instance age or sex would um affects the um kind of
00:11:22
the the the by market and value for this we used a course in process
00:11:27
a regression of the nice thing about uh this approach is that it's it kind of balances
00:11:33
the model fate and the uh so the century
00:11:38
how accurate the the model is and the um
00:11:48
so essentially the it it uh it buttons is the the model fit and the kind of
00:12:03
the complexity so essentially the uh so the sorry
00:12:12
i'm so the gospels aggression um is essentially described
00:12:23
by uh by a multi battered normal distribution are and what you see here so that's that's
00:12:29
kind of the the the movie about the cohen will t. barnum distribution with the it the means
00:12:34
and the and the combatants and here the f. and f. star f. is the observed data
00:12:42
so the training that essentially and that's part is the the the test data or the data where
00:12:47
we where we don't have a measures and now the nice thing is that we can condition
00:12:52
does this distribution by the by the observation so essentially we can we can uh i get to
00:13:01
the condition distribution of have started this would be that that and then none of the data
00:13:07
uh given the the observed a topless this this
00:13:12
um which will do that normal distribution and wasn't
00:13:16
that the the next thing he is again as i said before we don't have to specify
00:13:21
a kind of a very accurate basis function but we can actually have quite generic so this is
00:13:26
a very generic a kind of covariance function which is uh it does um regular basis function
00:13:34
which just kind of a looks at at similarity button has also some uh the new covariance and
00:13:41
and independent noise components are because we assume that uh every measurement has depends
00:13:46
noise and then we can simply uh minimise this also does does a loss function
00:13:53
and estimate the hyper parameters of these uh function here so and then our
00:14:02
assumption was that when we know only really correct we would get a better performance
00:14:08
uh classification of patients versus versus uh controls then when we would just um
00:14:16
correct the the by marcus with only a linear a regression but it turned out
00:14:24
this is shown in the on the on the op or uh upper i'm able
00:14:29
the the performance was pretty much put pretty similar so here a rock grey matter
00:14:37
means that is the features like without correction and then a k. l. r.
00:14:42
and b. bit conley no aggression um energy plot would be to get because impossible
00:14:47
question and you see that the performance in different uh tasks so this is a
00:14:53
onto the diseases that set before these are a patient with uh with early stage
00:14:58
of of dementia and like the bottom line um is that these did not
00:15:03
increase the uh the purple performance where it was helpful however was when we had
00:15:11
like a hypothetical scenario where the where
00:15:16
we would correct for the site where people
00:15:19
would be uh quite so because there's some fruit and some scandal by it so
00:15:23
uh scanners acquire paid up slightly different visited there there are systematic them scanner biases and
00:15:29
in these situations the the the court reporters
00:15:33
aggression which had this realisation a property performed
00:15:39
a substantially better uh then d. d. in your own regression so with and uh kind
00:15:52
of put these these kind of these these basic methods together and participated in a challenge
00:15:58
uh and we score uh on the on the third place so the the the first two
00:16:03
places where the like the same group with lex like is it the different them out wouldn't
00:16:07
so in terms of every on the curve we perform without place on
00:16:11
the other hand in in terms of of a chore actually see we
00:16:15
were pretty much in the middle so we had a um or our
00:16:20
algorithm or classify was not uh well calibrated and uh since since then
00:16:26
is that like this is this is still no group that is still
00:16:31
open this this does that benchmark but no group um kind of submitted
00:16:36
a better results and how is all this from two thousand fourteen so
00:16:43
we didn't thought okay that's we are we have a competitive the couple competitive
00:16:49
method so we should try to apply a relatively blindly in um in
00:16:54
in a in a clinical and sample and this wasn't very successful so
00:16:59
what you see here is the like a um a and r. c.
00:17:04
course where the black line here is on a kind of of one that
00:17:11
training set from from put data so this was like that that's that's
00:17:15
typically the performance that you would read in in in the literature uh
00:17:19
regarding um in the specification um or presentation of them are imaging images
00:17:27
in in your direction so that kind of the literature performance uh and the
00:17:31
blue line here uh for different um tasks is the will be obtained
00:17:37
from the sample of the clinical routines of these are were subject that
00:17:41
were came into the triangle routine and then we're assessing that what so
00:17:46
that means that the application to clinic between compact is is not not
00:17:51
a straightforward um and in other uh like similar study we observe the
00:17:56
very very similar um mark calibration of the classifier well to add on
00:18:02
the curve was kind of reasonable but when you see the the cup
00:18:06
of point here uh is not optimal and also the confusion matrix you see
00:18:10
that uh the essentially the yeah the the colour possible not optimal although
00:18:16
the classification i sent or the local decision value to our area under
00:18:21
the curve it's it's quite reasonable so the what we then thought is
00:18:30
that is that we cannot applied directly to the clinic um so we then
00:18:35
went on to to propose a kind of less of a black box
00:18:41
approach but to uh to visualise the that the features or the did
00:18:48
more or less the deviation off the different features across the brain uh
00:18:53
to the to the clinicians so since it is nothing else and then um
00:18:57
kind of the standardised deviation of the of the of the vocal
00:19:03
value in terms of grey matter and and a so called cerebral spinal
00:19:06
fluid as essentially what what you would what you see here is
00:19:10
when it's read then that means that this subject has um more like
00:19:16
and bring liquid in these regions and when it's blue then more than expected i mean based on old age and sex and when is
00:19:23
brewed and that means that the subject has less grey matter in
00:19:27
that a particular region then expected based on uh h. and and sex
00:19:32
so this we we expected that this would help
00:19:35
clinicians to um make better reading or better um
00:19:42
assessment of atrophy in a market so we tested that so we don't go to a web platform
00:19:49
where as austin that's what platform where people could or
00:19:53
the the the raters um had uh the option to kind
00:19:57
of for for ten key regions to assess whether to the
00:20:02
subject has either no atrophy like me into me that of
00:20:06
you or a pathological atrophy and we showed uh like in
00:20:11
in random order um either with or without the the colour
00:20:16
overlay so there were nine raters that involved um and we
00:20:22
evaluated like speed accuracy um consistency consistency means some of the
00:20:28
images were shown twice and then it was assessed like how how how many times the same rate the rate at the
00:20:34
same uh for for for the image um and then uh we also look at the agreement with uh some reference or
00:20:42
some reverence meetings and the the the four cheesy main messages
00:20:48
there or main findings of this study was that the the
00:20:53
display out of that of these core which is kind of addition information had no impact on the average height because we
00:21:01
would we expect that it could have an effect because either it's faster because they are totally guided by the like
00:21:06
but they're shiny colours or it's slower because they process on one hand the image without the colours and finish with the
00:21:12
calls but turned out there was no no difference um the
00:21:18
also the discourse david not in improve the the agreement with
00:21:24
the reference reading and also or or like the only thing
00:21:30
that kind of improved uh was that to the consistency within
00:21:36
the rate was improved so that means that the the by guided by the colours the readers tended to more often rape
00:21:44
exactly same thing on the same page and then another thing that is shown here on on these these graphs uh is
00:21:50
that essentially that the when we then took these these deviations
00:21:57
per region average uh and compare it to the to the
00:22:01
readings of the of the of the manual of the raters then this was pretty much a day that was high agreement
00:22:06
so that means that the computers actually look that it's very similar things this for a topic
00:22:15
in the in my thesis was about the image segmentation so for this we use the uh
00:22:21
keep learning and we kind of yeah we we uh so essentially my condition here was the
00:22:29
extension of the um to the unit that was developed um by all of on the better
00:22:34
but the reason was to stand is um the two three d.'s essentially using more efficient libraries
00:22:39
to uh to perform these uh these these conclusions and in three d. because otherwise one would uh
00:22:45
run into memory um limitations and then experimented with with uh uh with different uh like segmentation
00:22:53
costs of use uh these nations here i've been here to issue a segmentation and and also here
00:23:00
like other like some regions of the contents rhyme cortex and uh we also experiment with a or
00:23:10
yeah let's say the here just um like a very quick uh overview of of what the network
00:23:17
what that would ah so it's it's a fully convolutional network so that means it's computationally efficient because
00:23:23
it can um leverage or not not too we
00:23:28
don't don't um computation so it's competition efficient and also
00:23:31
it hasn't these so that the data flow is like from left to right and then on the
00:23:37
left side from up up to down and on the right side isn't from down to up and
00:23:43
these display these levels here so d. zero you zero d. one and d. two with that rock
00:23:48
are at different resolutions so in the top level it's at full resolution let's say one millimetre uh approximately
00:23:55
uh and then here every time in these putting latest uh that was using this is a off
00:24:02
um until this this level and then here it's the opposite so here we use uh of conclusions to
00:24:08
increase a delusion every step and then this means that this network at the ready to large uh
00:24:13
spatial context so for instance to predict the days of like these uh across here in the middle all
00:24:21
the boxes actually in three d. not only to d. that are in the um in the in
00:24:28
the green ones are in the right yeah about in the right um a square are counted as bizarre
00:24:37
as such i think it's seven fifty thousand have boxes that are considered adjust to predict this box but
00:24:43
still production time of one volume is less than a minute and also this we when we submitted for
00:24:53
um the result for benchmark in a digit segmentation we
00:24:56
scored a first place um at that time um so these
00:25:03
are just some some results that a kind of a century show that we perform level of state of the art
00:25:09
as good as one can compare for from the space that because of course the uh we have no exact a reference
00:25:16
data to compare with but uh these are kind of considered
00:25:20
um good numbers um we didn't this for for uh uh
00:25:25
other regions as well and essentially the the messages that i would be performed like
00:25:31
when you compare with literature on on a good level and he is just that
00:25:36
quite the vibration of the differences so what does it mean these are randomly uh
00:25:42
shows and um examples uh and you see the um like the green line the
00:25:49
the technical standard and and the the red line is like the productions not like green green lines where
00:25:55
calls on that and production overlap and red is the
00:26:00
you um where the where the uh the automated production
00:26:06
kind of over and over estimated the volume so it's like what everybody calls similarly like because the the
00:26:16
the the goal is not to to just measure the volume of of these these different reasons but it's more
00:26:21
about um assisting diagnoses and here we show that um the so to and so
00:26:29
when you look at this here these shows the air and look at the curve
00:26:33
of of the band of education of controls versus versus the mental patients and at
00:26:41
different stages and like the the the solid lines are for the for the minor segmentation
00:26:47
dashed lines so he printed or for the uh for the automatic to um and
00:26:53
is pretty much a like the same that carries the same penalty value um obviously
00:26:58
these point call it a strong a similar things we did she i would rather
00:27:02
be by mark is just to kind of to prove the point that the the like
00:27:08
this tool is has some potential so just as a wrap up for the regarding
00:27:17
the did the p. d. project uh it's so again three that were sweet kind
00:27:24
of main uh topics that i uh worked on doing the fees is um so
00:27:30
one is that image classification in we found that's actually meeting is kind of more reliable
00:27:34
and more useful for the clinic because we function rather come other uh difficulties uh
00:27:41
then that assisted assisted reading of atrophy it's uh at least we didn't expect the
00:27:48
results that we that we got uh and uh it is also kind of had
00:27:53
some strategic uh um impact on one kind of on the how the the memory clinic
00:28:00
um convey their their their um their patient care as
00:28:04
action on so and then in image segmentation they're just
00:28:09
essentially the the that that take a message is that
00:28:12
the these these days these people earning full convolutional networks
00:28:18
the unit albeit others uh are quite competent in these relatively simple a
00:28:24
segmentation tasks so for a second part which is about it's a currents kind
00:28:37
of current projects uh again in the in the in the field of
00:28:40
them are imaging uh it's i had these three uh topics in mind a
00:28:47
huge time restrictions i will skip the the first topic which is uh
00:28:54
about production production of survival length um in like my patience and i will
00:29:00
go directly to the to a a more challenging as segmentation tasks which
00:29:08
which is about um cerebral um michael bits on like the example that the
00:29:13
show before for instance from the white middle aged which are typically kind
00:29:18
of by by little radiologist delineated by simple modalities what century they take one
00:29:27
image contrast one image type and that's that's enough to to do that
00:29:33
to the phone segmentation and the the readers are relatively large um so that
00:29:38
means like hundreds or even thousands of boxes whereas in server my probably
00:29:43
it's a typical asians are one two three you of lesson five bottles and
00:29:49
it's not possible to find them uh on a single or like confirmed
00:29:55
that type and confirm them single um but that is so that means that
00:29:58
to be sufficient sensitive in specific we need multiple modalities and this leads to
00:30:02
this relatively complicated kind of flowchart that's the kind of the the way the radiologists
00:30:08
kind of find these these lesions and that you can see is relatively uh i'm
00:30:13
elaborated um and it needs a here are like different uh advanced uh imaging modalities
00:30:20
and i will show a just a example of this or that so here you see the red arrow is that
00:30:27
it's true region and the the the yellow arrow is the is the kind of so called the and unique so
00:30:33
essentially when you look just at this s. w. i. sequence that's also that's always the same brain but just a
00:30:39
different contrast you see that in both cases you have like a a dark spot here but then on the on
00:30:47
the other on the on the so called choose sam so i'm kind of excessive susceptibility uh image you see a
00:30:55
pat this part here is is dark but he also
00:30:59
you would see other like bright spots which confusing so in
00:31:03
order to to uh like to really get the two lesions types of the we also i'm into that so called
00:31:10
um and people's it's um it's necessary to to access at least three models which makes everything um complicate and again
00:31:18
uh as it like that's that's more or less like i'm not one or two or maybe three
00:31:23
text so very small uh visions um and you're just there other so that i i mentioned to
00:31:32
different types of your degree this is just another type of mission um so so to have this
00:31:37
multi class a classification is very small lesions uh we need it's it's not a lot of mental effort
00:31:43
although there's a evidence that is clinically relevant so the uh the result that we got our
00:31:51
m. so here maybe to say typically in segmentation tasks at the at the appalachian criteria some
00:31:58
overlap some measure of overlap like bass confusion here this will make much sense to to look
00:32:03
at uh at the overlap because like when it reaches one vauxhall and then uh to the
00:32:09
machine the text like the neighbouring box let's well then it's already like half half of the
00:32:15
of accuracy so we looked um on so called detection rates essentially we uh we did a
00:32:21
clever component labelling to find all the deletions and then looked looked around deletions whether there was
00:32:28
a what are the algorithm detected lesion within this uh uh ray just of of two millimetres
00:32:35
uh and then we we got like a reasonable uh sensitivity to what you see
00:32:39
here is like the server might with that one type of religion i'm deposits give it
00:32:44
all the time at regions uh and and different cat combinations of of of uh of
00:32:50
modalities and the the um kind of the just the sense to do was was quite
00:32:57
high but uh average position was like depending on the on the on tightly a comparatively
00:33:03
low um on the other hand in this is implemented in some some seem automatic i
00:33:07
wouldn't it could still be used as a as as as the help um and it's
00:33:13
a consistent with with other uh results um with these quite difficult segmentation past which now
00:33:22
um show that the more difficult task also for the planning a method
00:33:26
or not that's straight forward to sort it so the very quick another uh
00:33:35
current what which is some very close to my heart uh because
00:33:39
it's very close to to to a clinical um application where we use
00:33:43
deep learning to on yet to to to measure the the the volume
00:33:50
of the left option to this is hard to imaging not not brain but
00:33:53
very similar um concepts apply and and here the idea is here to to
00:34:01
get a t. t. two two d. images from from the heart and then
00:34:06
approximating the the volume of the of the heart um and this is done does dismantle procedure obviously
00:34:14
and this project we started from the now from the from the proposition that we said okay we
00:34:21
we do not want just to assist the the radio just but we wanted to do it in
00:34:25
a in a non disruptive weight was actually we said we want exactly the same process so now
00:34:30
additional colouring or something else but we um have exactly the uh the same process so
00:34:36
so we just which is this we uh kind of know to to the to the mantle stepped other done which include
00:34:45
image classification so there's a kind seems actually because the heart is beating uh so there's a bit high seems of images so
00:34:51
uh and the first step is to to detect so we specific faces in the in
00:34:56
the heart cycle so that you could um it's kind of an image detection or intoxication rather
00:35:03
uh then like six segmenting the the region of interest again a
00:35:09
manual process and then detecting landmarks on the on these images and then
00:35:16
measuring the some battle so you can conceive here so this this is
00:35:21
actually um and that and uh like approximated the fed the volume and
00:35:28
what we did we automated all of the steps with pretty much the
00:35:32
same base base a deep learning architecture and so does up does not
00:35:38
probably get as it is a kind of preliminary results but essentially a
00:35:42
we compare then the performance in that case uh of the of the overlap
00:35:46
on the on left side you see the the vibration set uh and on
00:35:50
the right side side the the the test set so we which was untouched
00:35:55
uh until until the end of the final model and uh the the the the basic message that
00:36:01
that in practised is pretty much the uh the or the performances in practised the same so you see
00:36:11
yeah for the low compared sewed with the blue is the blue is the sole rater one versus machine
00:36:20
the orange is ready to put his machine and then green would be to to break there's um compare
00:36:27
and also the like to do that but the nice thing about this is that
00:36:31
we could uh straightforward extend these two to three d. image stacks
00:36:36
without effort and have like even the more accurate uh estimation of the volume and this is like an interface
00:36:45
that is actually implemented at university of possible uh where
00:36:48
they now can get based meditation can start with this
00:36:54
created by the machines all ready the machine or a segment that's the brain that sorta cemented the left actually
00:37:01
uh put the landmarks here estimated the the time eaters
00:37:05
and then the the condition is still in full control
00:37:09
and still in charge and can change uh the thing or or leave it as is for a soul uh so
00:37:21
how much time to to them it's okay good good so solely now i've finally come to the to the um
00:37:30
might kind of positive some some basics of of uh
00:37:36
got why or in what direction i would like to
00:37:40
go into to next uh two years because so far what are the the the form until now was kind of
00:37:46
uh more or less guided by uh by kind of like to to the motivation
00:37:52
to to do with feces and to kind of too but just have uh some collaborations
00:37:57
um and for the for the u. s. postal postal mobility
00:38:02
a project that was kind of the first time i kind of
00:38:04
more more deeply thought about a possible approaches to some problems and
00:38:10
what would present now is not not the solution obviously but kind of the some background that i uh
00:38:19
collect that uh on do preparation of the of the of the project
00:38:24
so if you look at the or how i imagine the typical typical uh
00:38:35
landscape of of um of patients and and and and and and or or
00:38:41
just of the population is there is a yeah let's say hell see how
00:38:46
to control so group of how the controls and then there are different
00:38:51
manifestation of the same or different diseases but importantly these are like not the
00:39:00
same and also there's some some staging involved but with approaches the show before
00:39:06
with this image classification the typical assumption is that there's just two groups sort
00:39:14
um and now i'll kind of shortly kind of go through the some some
00:39:21
ideas that are out there already to to to look at these these heterogeneity um
00:39:27
and also also staging so what i mean by staging is that let's say if so this
00:39:36
is the has a control group um and then so on one hand they can for instance they
00:39:43
they get they get older so they go from from the stage to the to this page
00:39:48
but then they can develop different sub types of the same disease so it's just an expression of
00:39:55
how to the patients group could go from not only go through stages that are the
00:39:59
same for everyone but apparently outside subgroups um of staging okay and the kind of the
00:40:09
one of the driving force of this of this idea of identity in in your generation
00:40:13
comes from the so called a. t. and model eighty and stands for i mean all its power in your generation
00:40:20
so d. these are the three whole marks off alzheimer's dementia which is the
00:40:24
most widely um or like the most prevalent a new degenerative diseases um in
00:40:31
the elderly population and the thing is that by the by the book people's
00:40:38
all times they have i mean which houses a protein aggregates and the brain
00:40:43
i'm a new direction visible your direction but it turns out that that would
00:40:49
be only in if that would be would present a a t. n. n. but
00:40:53
only a small fraction of the population was all times actually has it and then
00:40:58
there are others that have either one a or two of the of these um
00:41:03
things but still some have symptoms some some some some some don't so essentially the the
00:41:10
very definition of of dimension is is uh uh is not very clear so and then
00:41:18
d. into said the the um the dysfunction
00:41:23
typically if you apply some like that is the
00:41:26
but the last function for the for for as the and it's only assumes that you have
00:41:31
one group and then the other group but then when the other group is like editing is then does this won't work um very well
00:41:38
um and this is a lot of friends it is something we recently did a so applying to this is called
00:41:44
um hi joe which essentially estimate the poly tool so instead of uh uh of just one separated by planes have like
00:41:51
multiple how high the plans for kind of some groups uh and this shows
00:41:55
for instance uh this is the population of a schizophrenia patients that this gives
00:42:00
edward genius um oh that is good idea
00:42:03
so some express different pathology hatch and then
00:42:09
uh then others and we want to find a similar things i mean in the match
00:42:16
uh and then other uh other methods uh include like that it took the classical uh kind of came in
00:42:23
clustering um and also this kind of does does does
00:42:28
consume computer vision from uh it's it gets into the
00:42:32
divisions called a clear and coherent point to drift algorithm
00:42:35
where the idea is that um all the different is
00:42:40
that patients come from from the has a population and
00:42:46
go there by a certain a transformation so that means that
00:42:50
that uh let's say it's fifty six year old male with certain um brain
00:42:55
pattern depending on his brain pattern will end up either here or here or here
00:43:00
and this is you in that case uh optimise our uh estimated with the
00:43:05
uh with the kind of abuse abrasion a model of these of these kind of
00:43:12
transformed google has controls um and the one issue
00:43:20
that we have um you imaging is the high dimensionality
00:43:24
uh and uh we we tackle these by by um
00:43:32
apply so called um no neck it nonnative may differ translation
00:43:36
uh which gives us essentially a clustering features and the clustering in subjects at the same time
00:43:43
so the um the the album up approximates that the
00:43:49
data also eight i would like to let's say a thousand
00:43:52
subjects and one million um boxes values uh with with the
00:43:58
um majors majors multiplication depending on constraints that one puts on
00:44:04
um on h. this can even be as simple as as um k. means clustering but uh with like
00:44:11
this budget constraints and so one one can get kind
00:44:14
of biologically biologically um let's say uh interesting um patents
00:44:22
and here currently we are looking about uh on the
00:44:25
on the stability of these decomposition so essentially um in
00:44:29
a it was but across the nation um manner we
00:44:35
we estimate these these discussions and then we do any
00:44:39
um i classifications with one of the uh the previous albums the hydra algorithm
00:44:43
and then we look at the consistency daddy measure in in so called um m.
00:44:47
a. i. r. so i'll just let them run index and the higher the better
00:44:51
i'm in principle uh and then we we see that when you have only little components
00:44:57
uh this gives like an over simplistic um approximation of grouping which in some cases even just
00:45:03
reflects a sex or even site so that just all the
00:45:07
subject we're scan on specific site are put in the same group
00:45:11
so that means that also here there's a lot lot to um to do with the interesting part is like that
00:45:17
uh and isn't that that is that like with more than twenty components
00:45:22
and so i will yeah so the these are like again very similar
00:45:34
uh on the similar but approaches to to tackle um heterogeneity also here
00:45:41
with a with a with a more like character clustering algorithm um that
00:45:47
also identifies different a sub types of of the mentioned that case
00:45:51
and this is what i mentioned uh before this
00:45:56
is an approach to uh to estimate for every subject
00:46:01
as staging and the sub typing at the same time um and we are going to use or plan
00:46:05
to use these kind of approaches as kind of a baseline a method to and have to come to
00:46:13
an end uh to essentially a uh yeah so i will have to just get to the last slide um
00:46:25
to just to show the you kind of the the what i eventually plan uh
00:46:30
to do or not i but together with of course the grouping that have yeah
00:46:34
uh is to have um what you see here in in
00:46:38
the centre is a a depiction of of um a latent variable
00:46:44
that little girl laden progression score in that case certain measures so to to measures look does that the that the remote thoughts
00:46:53
uh then the to dash line which is a a population trajectory
00:47:00
uh and the grey line which would then be the that case be bit deceptive specific uh trajectory
00:47:06
um and what you can see here is that uh although the second
00:47:10
measure like was higher than than than than the first one the estimation
00:47:15
still kind of negative similar to to um to to really makes model
00:47:21
uh where friends and eight would be on the on the x. axis
00:47:24
and then the the subject pacific slope would still be negative in that
00:47:30
case um but she are we obtain these these days progression score with
00:47:34
a with a generative model that essentially tries to map the subjects based
00:47:41
on the buyer marketing to aid in that case is shown a two dimensional
00:47:46
uh like the space that shows uh what what uh mm laden profile is is is the most
00:47:54
most likely and also you from here to here how it it
00:47:57
can evolve over time and with this uh one can then estimate the
00:48:02
future uh by market profile and then eventually uh linkage to some
00:48:11
other biological manifestation such as um clinical symptoms in order to uh to
00:48:16
kind of the closed loop and kind of make the case for
00:48:21
the for the latent variable presentation which there is no way to to
00:48:27
say what is the writer presentation but uh the goal is to
00:48:32
say whether the representation makes sense in terms of linking it to the
00:48:37
uh to to behavioural and then function characterisation and yeah with
Recommended talks
Finding delphi in medical imaging
Eldad Elnekave, CMO, Zebra MED
June 11, 2015 · 12:24 p.m.
178 views
Concept Measures to Explain Deep Learning Predictions in Medical Imaging
Mara Graziani, HES-SO Valais-Wallis
May 3, 2019 · 10:32 a.m.
458 views