0:00
Hi, yes, I'm mentioning I'm Marta Czapranska.
0:02
I'm a Senior Scientist at Spatial Biology in Concept Life Sciences.
0:07
And today I would like to talk to you about creative solution to scientific and technical challenges in digital pathology.
0:13
I would like to talk about two of those.
0:15
One, it's going to be a bit more scientific, another one is going to be a bit more technical.
0:20
But before I get into this, I would like to tell you a little bit about the company.
0:24
Yeah.
0:25
So Concept Life Sciences is a contract research organisation that specialises in delivering concept to the clinic, a concept to the clinic.
0:37
And we basically are able to deliver any of the processes along this discovery process from biology experiments.
0:56
OK, from biological studies and chemistry to drug morphology and pharmacokinetics.
1:06
So, the company is based across several sites in UK and I'm part of a biology team which is located in Edinburgh, and this is my spatial biology group.
1:18
Chris here is here with me and he's delivering a talk tomorrow.
1:22
So please all welcome and join us tomorrow as well for Chris's talk.
1:29
But at spatial biology, we really pride ourselves with delivering up to date processes because we’re continually assessing and upgrading our offering and catching up on the latest technology. We deliver everything under one roof from sourcing the tissue to image analysis and we're able to identify proteins, mRNA and viruses and deliver studies vertically and anything that could answer any scientific questions as well.
2:00
Most of us come from academic background and you're really interested in upgrading our knowledge constantly and catching up with latest scientific updates as well.
2:15
Today I’d like to tell you a little bit about image analysis services.
2:20
And we use Visiopharm software to conduct those.
2:24
Most of our studies are pretty standard from ranging from tissue segmentation like assessing tumour necrosis and stroma.
2:34
This is just one example of those, as well as measuring and phenotyping and so on.
2:40
However, what I'd like to really highlight that we really specialised in delivering projects that are quite bespoke, and we really take a lot of effort into answering specific client projects that could be sometimes a little bit more challenging than the standard things.
2:54
And one of those is fibre alignment.
2:58
So quite recently we had a client that approached us and wanted to know whether particular drug effects collagen fibre orientation in upper dermis.
3:08
So how this really came about is the client conducted a study when they wanted to initially assess area of collagen in human skin tissue.
3:19
And upon this assessment, pathologists noticed that in this particular disease state which caused lesions, the particular fibre orientation was very peculiar.
3:30
And he wanted to know whether that makes an effect, like the drugs make any effect in that sense.
3:37
So it was quite a preliminary study as a clinical study as well.
3:40
So I can't really give you full data, but what I wanted to talk about is how we actually came about to analyse this.
3:47
So does this, the fibre alignment even matter?
3:50
And the simple answer is yes, of course it does.
3:53
And I've got a couple of examples of this.
3:56
So you can see in a healthy tissue, the fibres are very much dishevelled, they're disoriented and become a lot more oriented in and aligned in a particular disease state such as cancer and fibrosis.
4:08
So this is cancer progression and as the cancer progresses, the fibre become a lot more aligned.
4:13
And this is the same true for fibrosis.
4:18
So there's a lot of methods already available for assessing collagen in literature.
4:25
Most of those techniques use second harmonic generation microscopy or software called FiberO.
4:31
But what we really wanted to do is basically use the samples that we already had and the tools that we already had for analysis rather than come up with another study design and spend more costs on it without even knowing whether this is worth it.
4:44
So how we did that?
4:46
So I'm showing in a large fragment of this tissue here.
4:50
And initially we used the chromacity red and polylinear local linear transformation in order to detect the features that we wanted to.
5:01
Then we did further object identification because basically we didn't really like any of those round little structures.
5:08
We really wanted those long fragments in order to be able to tell the orientation of this collagen fibres.
5:15
So we filtered this out and then in the end we kind of had to have a given measure towards which we'll be measuring the orientation of those fibres.
5:24
So we decided to use the epidermis here and basically draw a line as a reference point towards which we could oriented ourselves.
5:32
That gave us ability to detect vertical and horizontal fibres depending on the angle towards the reference.
5:41
So in our little test case where we're developing an APP, we had those vertical and horizontal fibres as you can see here, and they came about 70 to 30% ratio-ish.
5:53
But in a more realistic scenario, when you apply this to the entire tissue, we came across a few challenges.
6:00
One of those being, for example, the epidermis is obviously not exactly flat, and we had to segment those tissues in fair day in order to get good references where we could actually measure the vertical fibre orientation and more accurately.
6:14
And this came out to about 50% ratio.
6:18
So it wasn't exactly significant, and we didn't necessarily find a difference between the non-responders and the drug responders in this particular case.
6:27
We suggested further improvements to the client.
6:30
For example, as I mentioned before, this is a lesion state.
6:34
So those lesions, unless they have been across entire tissue, so segmenting that into smaller regions might have been a bit more accurate to actually tell us exactly what's going on.
6:43
But really it's also this was very low study number as well.
6:47
So boosting that up could actually bring some significance to this, if any.
6:52
However, the client was very happy with the approach and the results.
6:56
So and it was very meaningful to them in the end.
6:59
So we're very happy we've managed to work out a method to identify the fibre orientation without any expensive experimental design and specialistic microscopy and using the staining method that was already in place.
7:16
Speaking about staining, at Concept Life Sciences I’m part of the spatial biology team.
7:22
So this is very important to me.
7:23
We believe that good quality staining is extremely important.
7:27
Any steps of this concept of clinic pathway that I've shown you earlier, especially at the target identification and validation as well as lead identification optimization, we do believe the high quality staining generates the best foundation of good image analysis.
7:44
Obviously we can use techniques to filter our background and so on and so forth to bring out more staining.
7:49
But really it comes down to great meaningful staining in first place, and we take great care in any of those steps along the way.
7:59
Saying that, and this is another case we've had, the biology is being biology, and we do see variation in staining intensity across different samples.
8:09
This is a very peculiar case that we had.
8:12
We came across data set that was very varied across the intensity of different stainings.
8:18
And this particular panel consisted of fibroblast proliferation endothelial markers.
8:25
For the purpose of this, I would just call them Channel 1, Channel 2, Channel 3 because it's not really relevant for the message I'm trying to convey.
8:32
It's the method behind it is kind of more what I'm trying to show here.
8:37
But the aim of this was to correct characterization of proliferating cells in skin samples treated with specific compounds versus non treated controls.
8:49
So in this particular data set, we have found that some typical scenario we have sometimes this is example of different images from different samples.
9:01
And sometimes you get images that are very highly stained.
9:04
And this corresponds to the graph here on the scale.
9:08
Here you've got images that stained very low with the low intensity and images that stained with the high intensity.
9:14
And each one of those is a given channel and each row represents a different sample and then different image.
9:23
So quite often you get samples that stained with low intensity like this image here, and samples that are stained quite highly.
9:30
Like I don't know if I’ve got shaky hands, so I can't really point accurately, but you can see it here.
9:35
So for example, like this image.
9:37
But what we've seen in this data set is a huge variation between the intensity of different markers in a given image.
9:44
So for example, there's an image that has got high staining of two markers and a very low stain of the third one, which creates quite a lot of problems and it's quite hard to capture and analyse.
9:56
So we've managed to analyse this data set, but we were not quite happy with the approach.
10:02
It took us a long time, and we wanted to improve that further.
10:05
So this is an example of a low staining image and a high staining image.
10:10
And as you can see here, it's got very distinct pattern and it's quite easy to show you what I mean.
10:15
So in this case, in the high staining image, the background of this is very close in intensity to the lowest image here.
10:24
And we can definitely see this is true staining rather than some artificial that we might have wanted to filter out because of the staining pattern.
10:33
If I were to use the typical approach of intensity staining as a thresholding feature, I probably would have very hard time to analyse this as I did.
10:43
So what we did instead, we developed a deep learning APP and was able to detect the features of the given marker.
10:51
So it was it wasn't just based on the intensity of disdaining, but also of things like for example pattern of the staining, which allowed me to detect those with a high accuracy.
11:02
I then use the cell segmentation in order to identify particular cells.
11:06
So those are defined by the nucleus.
11:09
And then combining these two together, I was able to detect the cells that are expressing the given marker.
11:16
So in order to apply this to entire tissue and whole cohort of the samples, I used something called Phenoplex.
11:25
And Phenoplex has been developed by Visiopharm.
11:28
And Sunita, who's here in the audience might be able to tell you a lot more details about this, so feel free to go and ask her.
11:35
But it's a very effective, powerful tool, very simple to use.
11:41
And it's based on the intensity thresholding.
11:45
So we use that in a combination with our deep learning APPs in order to get the accurate selection of the markers that we wanted to.
11:54
So what I should have said before is those features are not just 0/1 type of score is either mark as present or not.
12:03
It's a probability range going from zero to 1.
12:06
Zero being marker is definitely not present, 100% not present and one being marker is 100% present.
12:15
So depending how well the APP is being trained, this comes with each pixel comes with a whole range of probability across the line.
12:24
So you can then use that to further select anything that might have been wrongly selected in first instance just by sliding those bars in and out.
12:34
I must say our deep learning APPs have been very well trained.
12:37
So the range that we've selected is pretty much across the whole sample.
12:40
So we're pretty happy with this.
12:42
But it's definitely easy to use and doesn't require any further knowledge about the marker presence or not.
12:53
As I said, we initially analysed this with our original method and that took us about a month.
13:01
With this deep learning capability combined with the Phenoplex, it was about a week.
13:07
So it was very effective.
13:10
Number of APPs that were initially generated with the original method was about 7, maybe more.
13:15
And then with this deep learning APP was 3.
13:18
So each corresponded to a given marker.
13:20
So it was very effective.
13:24
We don't have to do further adjustment.
13:25
As we say it's a variation across the whole sample.
13:28
So about 30% of images per given marker have to be further adjusted with the original method in order to get precise results.
13:36
With this it was about 10%.
13:39
But these are not the same thing.
13:41
So for the original method I had to use a lot of what we call post processing steps.
13:45
Those are basically further filtering out based on intensity and other factors depending what I've needed to do.
13:52
And this is very laborious.
13:54
It's very time consuming and they have to really make sure you're matching it with the given image takes a lot of time. With this deep learning method, as I said, it's a simple sliding this bar a little bit in and out.
14:06
Depending what I've decided, it might have not been selected properly in fancy instance. And the scalability of this approach some but very difficult.
14:18
I mean, I probably could have reused some of those APPs in further studies, but I would have been very, it was difficult in first place.
14:26
So into further studies might have been difficult as well.
14:30
With this, I can use any of those deep learning APPs and combine it with any other marker and be able to still use it for any different studies that I wanted to.
14:41
So in the end, I can build up a library of the deep learning arms detecting given markers and use them in any kind of combination, which is very effective.
14:50
So the bottom line, we're able to precisely select each marker and generate reliable data assessment.
14:59
In this particular study, we have found a slight increase in a proliferation group.
15:05
So this is a control group who hasn't been treated with any drug.
15:09
These are two different drugs that we use.
15:11
Well the client has used and wanted to assess, but this was a very low sample size.
15:20
So this study is now being further expanded, and this dataset might change depending on the results of the further studies.
15:29
But the bottom line is this method of analysis was easily transferred onto further work.
15:36
So just a quick conclusion, Concept Life Sciences offers a very wide range of biology services, including spatial biology and image analysis tailored to biomarker discovery and characterization.
15:48
And but by combining academic insight and deep learning, deep industry expertise, we support biomarker discovery and characterization in ways that directly inform drug development and translational research.
16:02
And our main focus is on delivering robust, reproducible and high precision spatial biology data.
16:10
I would very much like to thank my colleagues from the spatial biology team, Giulia, Lydia and Chris, as well as Justyna, Rhoanne and Hayley from Concept Life Sciences.
16:19
And very big thank you to Thomas and Oliver from Visio Farm for the constant support and help.
16:28
Thank you very much.