Tip:
Highlight text to annotate it
X
David Gutman: Okay. Well, I wanted to thank the organizing
committee for giving me the opportunity to present some of the work that me and a large
number of colleagues have been working at at Emory and a number of other sites across
the country.
So a lot of the talks that have been here have been very genomics focused and what our
group is particularly interested in is sort of an imaging-based view of cancer. So one
of my colleagues, Dr. Cooper, presented some of his work with glioblastomas and pathology
data. And what we're actually focusing on in this talk is looking at some of the radiology
data that is also available for a smaller set of the TCGA GBM cases, but we're hoping
to build that archive.
So, like many of the other groups, we're looking at correlations of certain features with outcome
and also with genetic profile. So again, what we're specifically focused on this talk is
radiology derived features. So as many of you know, glioblastoma is a grade 4 astrocytoma,
and is the most common form of brain tumors with a very poor survival and it was one of
the first tumor types to be included in the TCGA pilot project. So as most of you know,
mRNA data, copy number, DNA methylation data is available but it's not as obvious but the
portal but there's also neuroimaging data on a cancer imaging archive as well as whole
slide imaging so our group is mostly focusing on these two data types.
So again, the general methodology employed in our Asilica Center [spelled phonetically]
is to develop human and/or machine based assessments of imaging features. So in order to actually
do this, we have a large number of neuroradiologists that we've been working with, between six
and nine depending on the time of the year, that have actually been going through and
actually annotating these cases for different features. Now the question is, what are these
features that they're actually marking up? And so one of the things that there's a -- I'm
not going to go into detail, but there was a large amount of effort coming up with essentially
a standard vocabulary to describe what these brain tumors look like.
So similar to the pathology data, what we're interested in is looking for extra signal
that is embedded in the radiology data and the pathology data. Every case that we've
looked at has a diagnosis of GBM, but when you actually look at the images, similar to
everything else we've talked about, there's a huge amount of heterogeneity. And we want
to try to capture that information in a structured way. Particularly Adam Flanders and Carl Jaffe
[spelled phonetically] came up with a standardized set and there are lots of iterations of this,
but basically it's all of the things you could think of that would describe the tumor: Size,
location, as well as different imaging properties.
Now if anyone's curious what Vasari stands for -- it took me a while, I thought was an
acronym like everything else -- it turns out the data set that we initially used to validate
this feature set was called the Rembrandt data set. Apparently Rembrandt was -- the
biographer of Rembrandt's name was Vasari so they called it the Vasari feature set.
So don't try [spelled phonetically] to spend any time trying to get that out.
But one of the big points that we want to talk about in terms of heterogeneity is again,
since we're coming from an imaging focused view of the world, you can have a piece of
tissue that genetically is identical but it can have significantly different outcomes.
So you could imagine a small piece of tumor that is adjacent to the motor strip has a
significantly, honestly, shorter outcome then one that was in the frontal lobe because when
they actually go in and try to do the resection you essentially -- you have to you can't be
as aggressive about it otherwise you'll leave the patient paralyzed. So this is again other
metadata and other ways of looking at this rich data set that's available.
So again, I'm actually not a neuroradiologist by training, I'm actually a psychiatrist,
so I know -- I can look at these things, we can identify things and essential what the
radio -- the neuroradiologists have done is come up with kind of validated ratings of
different imaging properties. The ones I'm actually focusing on in this talk are things
that you -- kind of descriptive of the tumor bulk. So, one of the ones that is very kind
of easy to wrap your hands around is the percentage of necrosis. So essentially, what the radiologists
do is kind of in their mind they have a mental image of how big the tumor is and then they
can essentially segment. You know, they can say this tumor is highly necrotic and this
tumor is -- basically has no signs of necrosis. And they're trained to do this and these are
some of the things that they happen to be good at. And we had very good inter-rater
and inter-rater reliability when we were doing these sorts of things.
Now another -- there's other -- there's four or five features that I'm actually focusing
on. You can see here I actually went back. Basically what happens is, in order to actually
standardize this across raters they actually have a PDF feature guide that they go back
and look at it and we spent a lot of time getting the vocabulary right so that people
agreed what they were looking at; people agreed what the words meant so that we could actually
get a standard set of ratings and then we had people read the cases.
So, the necrosis one is one of the features. The other one that is going to wind up coming
up a lot is the proportion of enhancing tumor. Essentially, what is very common and one of
the key features, at least from the MRI data and the GBMs is contrast enhancement. What
that means is you give gadolinium contrast agent IV [spelled phonetically] and then essentially
parts that were not bright on the T1 image suddenly become bright. And that's particularly
interesting kind of from a genetic standpoint is it's associated with essentially a breakdown
of the blood brain barrier, micro-vascular hyperplasia and essentially kind of funny
looking blood vessels. So you can imagine that would be something that would be interesting
to look at.
Again, I don't want to spend too much time going into the tooling but the actual process
of acquiring this data was quite a monumental effort that was greatly aided by a number
of our collaborators at the NCI. So basically, the neuroradiologists were given 10 or 20
cases. They downloaded them from the cancer imaging archive and then essentially they
had this kind of radiology workstation where there's a little plug in, the little Vasari
plug in, where essentially they go through all of the different imaging modalities. And
many of you are probably not particularly familiar with kind of clinical neuroimaging,
but they normally get five to 10 different scans of the head, all with different types
of imaging parameters and the neuroradiologists are trained to use these different imaging
modalities to look at different features.
So essentially they download the images. They can look at them here. They can do some cross
-- they can do some very simple measurements and then they basically go through and look
for these different features like the one I mentioned is eloquent cortex. Is it in an
area that you really can't resect? And then you can imagine that would have differential
survival.
So for the data that I'm about to present, we had markups that were -- we had cases that
were read by at least three neuroradiologists for 72 patients. We now have about 125 patients
that have been read, but this was all done not in time for this talk. And for the ratings
that I'm going to present, we have three ratings per person but these are actually collapsed
down to a single rating. And so basically similar to these other analyses that have
been presented, we're essentially using these imaging characteristics as a probe to get
a better idea of kind of genetic and survival implications of these sorts of things.
So basically the first, the easiest analysis we did is basically we looked to see if there's
any correlation between the present -- if you have more of this feature what does it
do to your survival? And so in this case the feature that stood out the most was the more
contrast enhancement you have, the shorter your survival was. And sort of as we this
argument that this imaging data can be useful, we also started doing some kind of multivariant
regression where we took a standard clinical model which usually has age -- at least for
GBM, the typical model has age, gender and a performance scale which is sort of how well
the patient is doing at the time of surgery. And basically we try to see if adding additional
information from the imaging data would actually make you, basically give you a better model,
give you better predictability. And this is basically what we're showing here is. In this
case, when we did stepwise linear regression, the Karnofsky score was obviously highly significant,
but basically when we started dichotomizing this and saying having a little bit of contrast
enhancement versus a lot was again a significant predictor.
Now probably more -- and then I have to -- am obliged to show a Kaplan Meier [spelled phonetically]
survival curve and again this was significant. And this -- one of the things that's kind
of nice about this kind of qualitative assessments is -- I'll talk about kind of more sophisticated
ways to do this -- but these sort of clinical rules of thumb become very useful for neuroradiologists
to actually look at and kind of keep in their mind because they can do them relatively quickly.
So, and fortunately I don't have to introduce this concept but this idea of these molecular
subtypes that are based on mRNA expression has been introduced multiple times and one
of the driving things that as we started doing our more molecular analysis is we basically
asked, if we had the proneural, the neural, the classical and the mesenchymal subtypes
are there certain imaging derive features that are more common depending on your specific
molecular genotype? And answer is, the least in this analysis the mesenchymal type was
noted to have significantly lower rates of non-contrast enhancement compared to other
tumors.
Similarly, the proneural subtype which is -- has a lot of interest in that specific
subtype because of some survival differences was noted to have a large -- a small degree
of contrast enhancement. So essentially what this means is, this is the area that essentially
contrast enhances after gadolinium and basically what happens is they kind of in their mind
they say, this is X units. The entire tumor bulk in this case is basically this entire
area of abnormality. And again, this is what neuroradiologists are trained to detect so
these are the type of parameters we're pulling out as -- to use as our probe.
Now some of the other things -- again, this is just really to touch and highlight on this
concept, essentially, is, we actually looked at some of the mutation data. Now EGFR mutants,
we discussed that very recently. It turns out, of the 72 patients that we did these
markups on there was only, as of a couple months ago, the mutation status was available
only on 50 of them. So it's a slightly smaller subset. But basically we wanted to see were
there any imaging characteristics that defined patients who are likely to have likely EGFR
mutations, and obviously I'm showing it because there was. So basically, EGFR patients had
larger T -- had a larger tumor, essentially, or a larger area of tumor abnormality. And
interestingly, the TP53 mutants actually were smaller than the wild [spelled phonetically]
types. And as a correlate that means EGFR mutants were larger in general than the TP53
mutants.
So just sort of to conclude my talk, the main point I really wanted to make in this is that
imaging based features can provide important diagnostic information, even after accounting
for other clinical variables and as we start doing these genetically defined sub-classifications
and looking at things that predict survival, keeping in mind some of these other kind of
obvious clinical factors like location of the tumor and how that affects surgery and
treatment, it becomes important as we try to subtype these things.
Current qualitative work suggests genotypes may be associated with these imaging phenotypes
and basically this really sets the stage for future work. So as we said, we're increasing
the sample size so we can, going from 70 to 120 is obviously going to give us a lot more
power. Also, we're actually starting to move from ordinal assessments where there's these
sort of categories that are actually easy for the radiologists to assess to a continuous
based assessment because we think that's going to be a more sensitive probe, which the field's
called volumetrics. That's just more technically difficult but this training dataset that we
have actually allows us to validate our algorithms and make sure we're actually doing what we're
supposed to be doing.
And kind of -- this exactly parallels the work that Dr. Parvin [spelled phonetically]
and also Dr. Cooper and my group presented earlier with the pathology data. We can actually
go in there and describe an even richer feature set that are things that are not easy to quantify
for a neuroradiologist. But you can go in and describe texture, you can describe exact
shape and all sorts of these other multiparameter -- multi-parametric properties of kind of
what the tumor looks like. And we can start using deriving additional probes to basically
get a better idea of what is the -- having an EGFR 3 mutation, what does it -- does it
make your tumor look different? And going back and forth now -- that should have worked.
Can you advance to the next slide? Okay, yes. I'm about done.
So again, I'm just -- I'm just about done with my talk I just want to give an acknowledgement
-- this is -- a huge number of people have actually been involved in this. And that's
what's nice about this community; it's been very ad hoc. So there's been a number of people
from Emory, including Lee Cooper and Chad Holderin Scott [spelled phonetically], collaborators
from Thomas Jefferson, Henry Ford, SAIC Frederick, John Fryman and Justin Kirby, BU, Carl Jaffe
and the NCI has been invaluable for statistical help, UVA, Arifca Colon [spelled phonetically]
at Harvard, Northwestern. And then finally -- this sort of to make a pitch -- if you
are a TCGA contributing site and you happen to have radiology data, we would love to have
it. And if you could -- we have ways to anonymize it and do a lot of kind of the grunt work
to actually get your data out there and shared. And we think having this additional resource
will really inform a lot of the other work that we're doing. And that's it. Thank you.
[applause]
Male Speaker: That's really exciting to see those clinical
correlates. Thanks, David. I think we'll hold questions so we can have Elaine's talk at
this point.