Embed code
gonna make everyone so i'm feeling quite privilege here to show a very
technical or a work because when compared to do the presentation
of this morning a feed him much behind in terms of vision
and platforms there are four four rip releasable which uh research
but the message on to convey here on at least for this domain
of american image analysis is that you can have advancement does
but the only moment where you know if it works on that is really when you run
it on lots of of cases and and then you can do is rip releasable research
so what i'd like to to show of today is a medical image
analysis method for prices i met scene and more in particular
the three d. risk of iron spectrum of that that uh i would
for this cried and then show about france validation on the
small a set of a of a round uh the big concern human city
and then how we can we could build an open access ready mix what but found
that the doctors can use to further very dated with the with the on patients
so this is the outline i would first introduced the the motivation which
is to do non invasive personalise estimation of concept written success
and the whole the methods how to get there was text operators integration functions
then this validation on this might get a little customer that assets and then the platforms punching with these conclusions
some education's behind this work is that we know that if you have a c. t. image of of it you more
that the the structure of this uh do you want to see it is you
reflects the natures for instance you can see does out different uh i didn't consume us in c. t.
you have regions which are very dense these are regions are rather active cancer cells
are all the regions otherwise under under changes et cetera so this is
very uh had originally was it's not enough to in that we don't you
fry those regions but it's also important to understand the eco system
of the different regions because they are like seven michael habitats that are
uh working in completing which of the the beach with each other
and if we are able to understand heterogeneity then valuable to crack and divide the type of some of cancer
and also there for two to predict if you can if the treatment can success on
so this is uh what we like to do when he's walking this is more generally regrouped in under the terms of of radio mix
and what we want to do here in particular is to extract a
treat extra features which are related to the morphological properties of tissue
and uh to do that to predict response to a rigid therapy basically so they
process is like that's you you have that you more here the native country
we extract the set of features and then we hope for instance that
uh the the these patients that we'd have a treatment success for even came with the puppy
will populate a similar region of the features base which would help us basic it we don't divide them and say okay
justin based on the image than intensively uh being a good to say that this the treatment would work on that
so this activity surrogates uh the current uh person meeting based on
biopsies where you go with a needle you take a piece
of tissue which is quite critical if you're in the lines are in the brain and then you run very uh yea
along the agenda make analyses and you end up with with on with with the with not come
and also because based on the major have an overview of that you more okay
we don't have the molecular precision that we have an overview of the heterogeneity
so of course there are lots of walking does the main uh included going does
but they are back to main limitations that would like to highlight here is that the first one
is that there is no separate and then id analysis of the node components so basically they always
a average the the the the measures of the entire to more than the for you
you mixing like or have it that's another uh some limitation is that usually the size
of this vector here is usually larger than the number of points you have in
the space and we all know that is really leading us meaning to just went
over fit so this is to aspects we'd like to address in this work
so that's the decorator and and see how we can uh use like local measures
of of text you're here and for that we define the general uh
framework for texture which considers that n. e. g. dimensional uh takes genesis
approach can be characterised by a set of local operators g. n.
and also part also look out in the sense that their response so if
you apply to separate out what particular position x. x. zero hindering image
the response only depends on this a small box on the small special support here
do you have an operator like that you have an image like that thing and to apply the separate from this image to extract properties locally
so for instance uh if you do that with the now operate toss
you have like the compulsion you we'll go a conversion operators
and so basically come all this image uh to you you that's got product of this guy
who is is this image at every position you end up with a response map
which would highlight specific project properties for instance uh for for for that you know you not textual that you
so in this work we use a a one particular type of operator which as our it's way let's
and basically the these guys they they work a bit like a direction that image theory
that you said so you can see that because this is the collection of of
text operator was that gives you now defining for you so you need to play
by jeremy guy here in the numeric toss again you get there that is
that are like somewhat normalised but the vector of this uh and we get here so you have kind of or bus image directories
and so they have an intuitive but interpretation because for instance we take all the one you you get the slope to
the gradients of the transition between the pixels although to is the is the the the the curve it's up there
so if you look at them when you can hold it with this is a
something that is band pass it looks like that's we have basically three d.
with that second order very directly them next wise et cetera and um
actually there is nice property with them is that uh there's terrible which means that we act
actually uh can express and rotations of them within our condition of the bases sediment
and which help us to activate locally i line and those with that's for each position to make
sure that we are normalised orientation we are locally rotation valiant of for for the images because
we expect to be able to describe is look at properties within vines to edition
okay so now we define local operate us and we'd we'd like to have
the aggregate them over region basically instance that you more so we
we apply the operators to all position in the image and then we we get the response now
but in yeah we want to have one ms you for that you more not a set of of of response everywhere
so we need to aggregates those measures uh to estimate feature statistics and we need to do that or region of interest
that could be for instance the core of that you more all the margin of two more on anything else so
to do the segregation there sort options natural one is to do integration and the the
most the simplest example is with average would you do the average of each
operator responds uh over that you also get a set of features here
uh but there is a very strong limitation was average as as we know it
might give some signal so for instance i like to do straight something here
so imagine we have this text you're here which was clearly two distinct a visual events that's it
and then we have to operate also job just like secretary symmetric here they just look
at things of difference gets here so we can see the response maps over here
and now we will aggregates this is a response or different regions here so we can aggregate
over the the right the uh all the green ones over or over the whole image
and you could look at the feature space that this stands so if we look at the the average
response of of just operate or here on the x. axis and separate on the y. axes
we can see that all the green regions are very well uh regrouped same thing for the red ones
and when you average over the want entire image with the blue one end up or something in me the rejected
doesn't correspond to anything does your average this is this is not happening this image that's a big problem
so coming back to our program of of an aside to mars this is
the problem because if you average picture properties over these heterogeneous two more
then you end up with something that doesn't correspond to him sing biological essentially get your signal and and you cannot do anything with
so let's see another way here to drug regulation which is still based on summarising the
the response of the boat also over region but men using the the covariance matrix
so now it's a bit better because we keep for each position the uh the
cool actually visions of the of the features if you want so we have a set of operators and we
know they the covariance together that's that's a given point so this is what we will use here
and then you can for instance the vector rice them here to to obtain actually if you recognise your components metrics
of just to go per triangular because they're symmetric you end up with with a with a vector like that
now we need to take into account the facts does does mattresses
we cannot just measure that couldn't distances between these vectors
because they leave on the symmetric positive this and if you need a menu for no
so it should take a a naked in distance between them this is wrong it's a
bit imagine you have to drive a from now that one out for instance that
it that's useful to know the the couldn't distance between these two points because you
cannot fly between them you need to take the role to go there
so this is the same saying you needs to try don't put this money
for the company that couldn't distances between these two guys not helpful
so they are when using distance that that exist if you want to compute
the distance between these two koreans but excellent going to do this here
and what if we want to do more than computing distance but to compute scallop
products for instance to do a c. m. classify our our feature or anything
well uh it's also possible that but you have to uh look at me
uh approximate the properties of human before so basically you take one production point which is
uh uh between the two a manual so i see this fear is is the menu full of the covariance matrices
and this is a production point and then you can uh project all your points
on this plane tangent here with what we would call logarithm exponential metrics operations
that's the point on this point here so i won't go into details but you you need to know that we can go from this menu
forced to displaying and on this plane we can do uh and it's
got a product so we can compute as yams for instance
and you can have your decision functions that taking to a account
like one man and a covariance matrix and the train hype
then of course to stop products depend on the production points so and it's computed like that
alright so these other but that's that that we have that we have and let's see how it works on a on a i dunno customer
so what we wanted to do here is we have a ninety to run like
uh the customer from uh our colleagues that lets them for hospitals and clinics
for each of these uh cases we have uh the regions donated
to have the grass to mow volume should entire remote
the grandest capacity which is the room want excluding the what one and sorry
regions which are the most familiar news which are the red ones
so we will use the different regions recognition to see if it makes a difference and also we have the disease free survival
time so which means that we know that this is t. it was zero and then they started the register putrid
and then that detracts which patients have the relapse during this period
not will do in this war could just simply a binary thing so we will consider all the patients
that's how the relapse before twelve months he before one year or not okay so
this is our that binary amount that we want you to uh to predict
now we we estimated generalisation performance using the telephone 'cause edition
for each training fall we need to estimate this projection point for the menu full
of basically will use the minutes of the koreans of the training set
uh so that it's cool to everyone and you will of course train after as yams on this on this tension plane
for each test for that we will then compute the classification groups so these are the results
so we fixed some parameters of our framework we repeated the the experiment several times
we have a here quite and buttons that the sets because you have twenty three recurrence within the for twelve mourns next
to sixty nine or missions and these are the the results so that makes bandwidth what we have in the stables
i i is when we use the full framework so using the s. yams a trained on the menu fault
b. is when we use the covariance matrices but we compute okay then distances uh
between them since actually you can fly between uh the and one of the
and each line here is when you agree getting over is the ontario to
most of the g. t. v. sorted in red or did you
that's mainly focused on the f. one score because this isn't balanced so so uh this
is a better uh ms your for in buttons that asset if one score
and we can see that already just using the the fact that this covariance leave on the on the
the symmetry defeats the many fall we can already improve a little bit to the the performance
uh and especially when we're gonna get over the entire to go for some seventy six to eighty
that's such an also interesting to see what kind of results we gets when we
do would simply do average over the of that you or that you more
and then we we see that this is a big problem because here when you have a range of the
anti r. two more the g. t. v. uh we get the really big drop in terms of performance
and this is a big plus the case it should just to get them on the certain parts because those guys every how much uh news so it's
still okay to agree get an homage is regions of course but the prime
happens is when you do an average over over a hydrogen you see
so i just want to our bodily had happy so it seems to work as we want
but it's still only ninety two patients with a number features just is about the same
so we we don't know if this so this is all on your feet on
so this is why we are no building a a here an open access when you mix with by phone
corrected it talks like that so you can uh go to this website you can test it
and that then you for the doctor last they can just a big zip file with
all their passions and then they can have like a c. team age even pets
and and the uh the region of interest you can submits and it would process them in the batch months we
can go uh we with hundreds of patients you you choose a few parameters packet textures and things like that
then you can get results like uh in this easy file that you can plug
into your father it's a statistical software lex is p. s. s. or are
and also like basic ways to to miss your uh to use your statistics on the on the features
so we we started to use just got from now to further validate what we've
done and really to your hands to the doctors that they can ah
a pro that's of patients at this and see if if it works for for
the problem we already started to have a very interesting results uh on that
it's also nice for me because i finished my job i i we program that and then that's it's the
doctors that need to find it it it and and and and get some people sort of it is
okay so it's time to go to conclusion that button perspectives and so we were able to predict written failures
uh within the first and pull phones for isn't gonna go see of of eighty percent so this is
but useful for doctors they can choose if you want to use the the treatment or not um you've
seen that is covariance a free market elegance to agree getting it without really killing the information basically
and we have beats this up and that's us a web platform of l. that is really
available now publicly also for last evaluation so we are instilled have dogs because okay
some people named differently there five since it was we were still into this phase of
making it work for for most people but uh i think we're getting there
and and the switch or work then we really want to to buy those red you mix uh features
for for being using pick a routine so so this needs a lot of patience like
it's you have hundred features we would like to have thousand patients seem like
so thank you very much for for for attention

Share this talk: 


Conference program

Welcome
Sébastien Marcel, Senior Researcher, IDIAP, Director of the Swiss Center for Biometrics Research and Testing
24 March 2017 · 9:17 a.m.
Keynote - Reproducibility and Open Science @ EPFL
Pierre Vandergheynst, EPFL VP for Education
24 March 2017 · 9:20 a.m.
Q&A: Keynote - Reproducibility and Open Science @ EPFL
Pierre Vandergheynst, EPFL VP for Education
24 March 2017 · 9:54 a.m.
VISCERAL and Evaluation-as-a-Service
Henning Müller, prof. HES-SO Valais-Wallis (unité e-health)
24 March 2017 · 11:35 a.m.
Q&A - VISCERAL and Evaluation-as-a-Service
Henning Müller, prof. HES-SO Valais-Wallis (unité e-health)
24 March 2017 · 12:07 p.m.

Recommended talks

Entreprendre : c’est naviguer entre les actions commando et les actions sociales
Domenico Savoye , Directeur - Bitz & Savoye SA
1 April 2014 · 7:05 p.m.