Authors: Michael J Walsh (Speaker), David Mayerich, Suman Setty, Andre Kajdacsy-Balla, and Rohit Bhargava
Mid-Infrared (IR) spectroscopic imaging is an emerging approach to derive chemical images from tissues based on their inherent biochemistry. Histologic diagnosis is the gold standard for evaluating the presence and severity of most diseases. Current histopathological techniques use of panel of special stains and immunohistochemistry (IHC) to assess tissue architecture, determine cell types present and to classify disease. Here, we report on the evaluation of an automated means to accurate histologic recognition using IR spectroscopic imaging. This method does not need dyes or probes and dispenses with human input but relies on computational approaches to provide decisions. Recent advances in IR spectroscopic imaging for tissue histopathology with a focus on breast cancer diagnosis will be presented. IR imaging coupled with computational approaches has the potential be a powerful adjunct to current histopathological procedures, with the ability to take a single unstained tissue section and give decisions on the cell types present and to provide novel chemical information to the pathologist in a useful format.
So now I'd like to welcome Dr. Michael Walsh to give a presentation, he's currently a postdoctoral
fellow in the lab with Dr. Rohit Bhargava in the department of bioengineering here.
And today he'll be talking about chemical imaging for histopathology, an emerging route
for molecular and structural analysis of tissues.
Thank you.
Are you there?
Yep, I'm good.
Cool, thank you.
Cool, good afternoon.
First of all, I'd like to thank the organisers of this symposium for giving me the opportunity
to speak to you all today, it's a real pleasure to be here.
Didn't have to come very far, just four floors down.
So today I'd like to talk about the emergence of chemical imaging, so a new imaging modality
for histopathology with a particular focus on cancer pathology.
So if you're interested in looking at tissue, particularly biopsies, and you're interested
in looking at molecular and structural analysis, you'll typically go and use a stain.
This can be a special stain or immunohistochemical stain.
This is an example of a colon with some work underway, Rex Gaskins.
So what we have is a new imaging modality, or a series of imaging modalities, which we're
terming chemical imaging.
And this can give us similar information, but in a completely label-free approach, and
it's based essentially on the inherent biochemistry of the tissue.
So how does this work?
So one of these modalities we call, it's called FTIR, mid-infrared spectroscopy.
And the basis of this is that we use mid-infrared light, and different regions of the mid-infrared
are absorbed by different chemical bonds.
So for example, carbon oxygen, carbon nitrogen, and these chemical bonds form the base constituents
of important cellular biomolecules.
So for example, proteins, DNA, lipids, collagen, glycogen.
So the principle is we have our mid-infrared.
It passes through our tissue, and we get what we have is an absorbance spectrum, or
a biochemical cell fingerprint.
This is part of the spectrum.
And essentially, these different peaks correspond to different biomolecules.
So these are a measure of the chemical bonds.
They vibrate in different modes, and this will tell you what structure they're in, or
things like protein conformational changes.
And essentially, the basis for all I'm going to show today is that this signature, this
spectrum, will be altered between different cell types, or different disease states, or
whatever it is you're interested in.
Okay.
So what is FTIR?
Well, it's actually very similar to a visible microscope.
And in fact, you can do visible microscopy through this.
The difference is we flip a mirror, and instead of visible light, we use mid-infrared light.
And instead of using a CCD, we have an infrared-sensitive array detector.
So one of the biggest advances in this, and which has really led to this being an emerging
technology over the last few years, has been advances in these detectors.
They've become much, much bigger, and now we're actually on a scale where we are clinically
feasible to image within a few minutes.
So what do we get?
So we'll end up with an infrared data cube.
So you'll put your tissue on, you'll press the imaging button, and then you get this.
And this is a data cube, our infrared data cube.
So it's actually very similar to a typical visible image.
We have our X and Y.
There we go.
This is our X and Y.
This is our spatial dimension.
And then we have our Z dimension.
So in a visible image, this will typically be red, green, and blue.
In our image, we have all this chemical information.
We have all the chemistry.
So essentially, every pixel within an image will have the entire spectrum present.
So you'll have everything you need.
Okay.
So we're using this technology primarily towards tissue histopathology.
So by this, I mean being able to diagnose diseases, particularly with a focus on cancer
pathology.
So if you're suspected of having a cancer, you'll have a biopsy taken, and you'll have
typically the first stage will be a H&E stain.
Many of you are familiar with this.
And then alongside this, you may do a series of additional serial sections and additional
stains.
So these can be special stains, such as Mason's trichrome, or cell-type specific stains, such
as cytokeratin for epithelial cells, or vimentin for fibroblasts.
And then you may take another series of stains.
So, for example, in the case of breast cancer, things like ER, PR, and HER2-neu.
This is a very time-consuming, expensive, laborious process, which doesn't always work,
as any of you with immunohistochemical stains will be able to attest to.
So our goal is to integrate all this information and take it from a single unstained IR image.
So we can basically mine into the chemical information, for example, the collagen protein,
and use this for classification.
And I'll just generally tell you, I won't go into too much detail with the classification,
but I'll give you an idea of what it gives you.
So the first thing we do is we take our unstained biopsy sample.
The important thing here is this is completely compatible with the current clinical workflow
in pathology.
I work very closely with a call hospital adjacent to here, and especially with the pathology
department at the College of Medicine, the Department of Pathology in Chicago.
So we take our unstained biopsy sample.
We'll do our staining.
You do immunohistochemical stains.
Alongside this, we'll collect our infrared data cube.
As I said, we work very closely with a pathologist, and they'll give us information on this,
and maybe cell type, or maybe disease subtype, whatever project we're interested in.
And then we'll use some sort of computation to give us similar information.
So I'll give you a couple of examples in a moment of using Bayesian classifiers or
artificial neural networks.
So the goal is to try to reduce the burden on the pathologist and develop infrared imaging
to give the same information or potentially more information.
Okay, so this whole process actually involves many different elements, through to the theory,
instrumentation.
We do a lot of work building novel infrared imaging instruments and Raman imaging instruments,
data analysis, sampling.
And we have some posters tonight and tomorrow with some examples of these.
Okay, so I want to give you a couple of examples of what this can give you.
One of the things I've been working on is being able to histologically segment breast tissue
into its key components.
So this particular study was on eight components.
So epithelial cells, a number of stromal components, such as fibroblasts,
cancer-activated fibroblasts, pure collagen, blood necrosis, lymphocytes, mucin.
So the idea behind this is we take our infrared image,
I then take multiple special immunosuppressant strains,
and we design a training set to basically be able to predict what the different cell types are
in this on-strain tissue.
So the basis behind this is we'll build up our training set,
we'll extract the chemistry from all the different components.
These are the components here.
And we basically data mine and try to find features, spectral features,
which are chemical features, which are predictive of what component it is.
So a nice example here is this big orange spectrum here.
This is mucin.
So you tend to find that mucin is highly absorbing 1030, which is glycogen.
This is a glycoprotein.
So what we do, I won't bore you with all the classification steps,
but what we can do is we can mine into this and use this for predicting the components.
And this works very, very well.
We can get an average area under the ROC curve of 0.92 in training, 0.89 in validation.
But essentially the take-home message is we can get our infrared data cube,
take each pixel, run our classifier, look for these metrics we've identified,
and we'll segment it and predict what component it classifies to.
So this is a really nice example here where to identify all the different components
in this tissue, you'd need to take five stains, you'd have to sit down,
you'd have to do a process of deduction for what's staining, what's not staining.
What we can do is take our single unstained IR image, run the classifier,
and it will colour-code everything according to what is there.
So you can see your cancer-activated fibroblasts and your fibroblasts in the middle
and epithelial cells in green.
However, there's one big major limitation with this technology, is the spatial resolution.
If you go out and purchase an FTIR instrument, the maximum resolution you're going to get
is about five by five microns, which is fine for everything I've just shown you,
but there are certainly cell types and tissue components which are really important
for diagnosis or just for looking at disease progression,
which you can't see with some examples here.
So one of the things I've been working on during my postdoc has been advancing
high-resolution imaging approaches and applying these to tissue.
So this is the sort of standard, it's from prostate core that you'd get
when I was doing my PhD a few years ago, about 10 by 10 micron.
This has gone up to about six by six micron.
And now we have, we've been working on something called ATR,
which gives us an increased resolution by higher NA.
And this is now what we can get using our specially designed instruments in our labs,
and Rohith has a poster tonight.
Yep, it'll tell you the theory behind the instrument and how it works.
And we essentially published this in Nature Methods last year,
where we could get these dramatic improvements in spatial resolutions
by modifying our instruments.
And this was the first ever time that this was shown to be capable
with mid-IR instruments and certainly on tissue.
And this opened up a whole host of new components for us.
So this is an example here, this was published earlier this year,
where we can look at endothelial cells, we can chemically characterize them,
which simply weren't accessible before.
Now this is really important because this is a sort of conventional image,
this is a high-resolution image.
You can see here, this is in a breast tissue,
the dark regions here, the dark purple regions here,
they are epithelial cells and they have infiltrating intralobular stroma.
This is really important to visualise for early breast cancer detection.
Using our conventional techniques, you can't do this.
When you extract chemistry from these different components,
so epithelial in green versus intralobular stroma in purple,
you get a greater chemical contrast.
And this is really important because this is exactly what drives our classifiers.
So we're doing a lot of work to be able to do very, very accurate,
high-resolution segmentation of all the different components that we need.
Okay, just to wrap up.
One of the things that this does,
so I've shown how we can accurately segment our tissue based on the infrared images.
However, one of the problems is that we're throwing away a lot of information,
the staining is still very important,
there's a lot of information to be derived from that.
So I've been working with a Beckman fellow here,
who's a computer scientist,
on developing artificial neural networks to basically be able to replicate staining.
So what we do is we take a visible image,
this can be HD or trichrome, whatever you want,
and then we take our infrared image.
We then overlay them very, very carefully.
And we can train the neural network to identify contrast differences
in the visible image to chemical differences in the infrared image.
And essentially what this can do is we take our own staining section,
run our classifier, and it'll predict how it'll stain.
And this works really nicely.
So you can see this is our real HD on the right,
and this is our digitally reconstructed HD on the right.
We can do this for large sets of tissues very fast.
So what's really nice with this is because we're using chemicals and not,
because we're using the inherent biochemistry and not stains,
we're not subject to some of the limitations with staining.
So you can see here where many of you will have seen this before,
where the stain just doesn't reach the end of the tissue,
or something's gone wrong.
Because we're using the inherent biochemistry, this doesn't happen.
We've basically now coupled this with our new high-resolution infrared image
in our lab.
It's the only one in the world, actually, offline.
And we can do really beautiful images.
So we can reconstruct HD images, Mason's trichrome images,
even immunohistochemical stains.
We've done this for a whole host of things by Menton and CK, CD31, P63.
And what's really nice is this is just a single section.
So you can scan your tissue, click a button,
and you can do all these different stains.
If you have the stain and you have the IR image,
you can basically construct.
Everything we've tried to construct so far has worked really well.
Sorry, it kicks me off.
What makes this a good modality?
It's fast. It's potentially elevated.
It's completely not destructive.
You can stain the tissue afterwards if you wish.
From a clinical perspective, there's no expensive stains or antibodies.
It's very fast.
Importantly, it's amenable for quantification as well.
For example, if you want to look at inflammation,
you can run a lymphocyte classifier, click a button,
and it will give you a very accurate percentage
of how much lymphocytes are in your tissue.
What's also very exciting is it can potentially provide
new chemical information into disease progression.
So something I haven't talked about today,
but I'm looking at the advancement of pre-diabetic
and diabetic patients and transplant patients
and being able to look at glycation in a way
that's not been possible previously.
So with that, I'd like to say thank you very much.
As I said, we work in Roy Bargarver's group.
We have five posters, so don't harass the questions.
You can listen about skin and prostate and the HD instrument.
Sarah's got the poster on tissue engineering and using infrared.
So thank you very much.
Thank you.
Are there any questions for this speaker?
It's about how protein works.
I'm wondering, when you say protein,
that tends to be a sheath that's hidden in certain characters.
Is there any way that that can become more specific
than that on certain proteins?
That's a good question.
So you can detect, oh, you can detect alpha helical
being depleted.
So one thing you can do if you're interested in protein
is use spectral deconvolution, and you can actually pull out
all those different components.
It's really, really important.
It actually is often very, very useful for diagnosis,
certainly things like this.
In regard to single proteins, probably not,
just because it's more of a holistic protein level.
So one thing we've been working on is looking at ER positive
versus ER negative breast cancer and PR.
Early results look promising, but there's issues
with epithelial heterogeneity.
You're not necessarily going to detect a single protein,
but you'll detect the downstream effects.
If you're ER positive or ER negative, there's a whole machinery
which changes, and that has quite dramatic chemical changes.
So, yeah, it's a work in progress.
We have some really good results, but it's not working as well
in validation as we'd hope.
So where's your space?
Could you tell us a little bit about your APR system?
Is this, like, do you use a wet sample there,
or what material is APR?
Yeah, it's actually a germanium crystal.
It's a 500 by 500 micron germanium crystal.
We actually do it on just dry tissue,
so this is paraffin-embedded, well, deparaffinized tissue.
You can do it on wet tissue.
So typically in medium phase, you can't do in vivo work
because water's absorbing, absorbing the medium phase.
But if you do an ATR-based approach, you can do,
potentially do in vivo imaging.
Yeah, so basically, it's a high refractive index,
so you have a higher NA, and you can then get high resolution, too.
So we get about a 1.56 by 1.56 micron spatial resolution.
Thanks.
It was a very nice talk.
Thanks.
So I just didn't quite understand.
So this is an actual hyperspectral technique?
Are you getting a pixel-by-pixel spectrum,
or is it just kind of an image of the image?
No, you get, it's a hyperspectral image.
You get all the chemical information, yeah.
And so are you able to separate out,
like is there usually a lot of overlap between signals of,
you know, different stains or things like that,
or do you have to do deconvolution?
As in spatial deconvolution or spectral?
The spectral deconvolution.
Yeah, so we don't typically do spectral deconvolution.
For this point, we're not too interested in where it's coming from,
from a protein level, for example.
The deconvolution's been interesting.
It's been tricky.
There's been a lot of sort of, it's probably primarily
related to the protein.
People who are more interested in those alpha helical
would be depleted.
But no, we haven't done a lot of that.
I saw a presentation about six months ago,
and yeah, basically they give very different data depending
on what deconvolution technique we use.
So we've been looking at spatial deconvolution.
I don't think I'll post on that today, but spectral deconvolution,
no, we haven't done a lot of them.
There's potentially a lot of information.
You can do things like derivatives,
and that will give you some of that information as well.
Thank you for your talk.
So I know that this work is definitely in advance.
We're replacing the time intensity and the cost for staying.
Do you still require a human in the loop for reading the pathology data
and making the cancer diagnosis?
Are you working on making your classification
predictive for cancer screening?
Yeah, that's a good question.
And that's something we've been working very closely with pathologists
with.
I think realistically we're never going to replace a pathologist.
It's just too complicated.
I think there's a lot of things we can do to help guide the pathologist
towards the final decision.
Certainly things like stains.
You have a limited sample.
I really wish I could do a CK stain for metastasis or whatever.
This could be very useful, but we're a long way away
from giving diagnoses.
It's just too complex, I think.
Things like sarcomas, I think we can have a big impact on.
That's something I'm interested in in the future where we can give
diagnosis of cell type of origin.
That will help.
It will still need a pathologist.
That's a good question because I think a lot of what we're doing
has to be useful to the pathologist.
A lot of stuff where it's just not useful.
So do they want to classify or do they want a stainless staining
type approach?
So yeah, that's still a work in progress.
That's going to be very important.