0:00
And I came up with the title of this presentation after seeing that all these rooms were named after famous actor.
0:06
So I thought the movie theme would be interesting.
0:09
I wanted to share with you some insights I've had over the past years, both working in a role as a diagnostics company, making diagnostic antibodies, immunohistochemistry kits for important biomarkers.
0:24
Later, when I worked with AstraZeneca, I was actually and looking at this from the other perspective, how to launch drug therapies with good diagnostic biomarkers for precision medicine.
0:37
And now lately for the past four years, I'm active in digital pathology.
0:41
So I have a quite broad view and I wanted to highlight some common conceptions, misconceptions also about immunohistochemistry, the work of pathologists and see there's a few in the room and also give you some future prospects on AI analysis and how specifically AI analysis can be used to train pathologist and how to make it more easily accessible for routine use.
1:09
So the first thing I did, I want, I asked an AI algorithm to make this nice graphic for specifically for this presentation.
1:19
And I noticed it's not very good at spelling.
1:24
It also must have some ethics behind because it removed the cigarette from the hand of Peter Sellers, and it also even removed the nuclear bomb icon.
1:34
But it did add some, I think very ominous look for this typical pathologist.
1:41
But again, it's trying to play with AI is something I would recommend to everyone, specifically in the field of diagnostic pathology.
1:51
So what is Pathomation?
1:52
Pathomation is an IMS company.
1:56
So we make software to view digital slides.
2:00
We can handle any kind of image format.
2:02
We have diagnostic workflows, we have research workflows, and we have educational workflows.
2:08
So different applications across the board.
2:11
The company was started in 2012 by two pathologists and a bioinformatician.
2:19
We have had, let's say, an organic gradual growth.
2:22
Today we have 17 people on the team.
2:25
We are based in Antwerp and Belgium, and we have, amongst other things now CE IVD certification for the software.
2:33
We are a medical device manufacturer, and we have basically a global presence.
2:37
So there's probably some customers in this room at this very moment.
2:44
What about immunohistochemistry?
2:46
So if you ask oncologists, molecular biologists, you know, to badmouth immunohistochemistry, they will typically say these things.
2:55
It's not very reproducible.
2:57
Yeah, it's only semi quantitative, which is correct.
3:01
But that doesn't mean it cannot be used as a good biomarker.
3:07
And what you very often hear, yeah, it's just a human eye.
3:10
So pathologists are eyeballing this.
3:12
How accurate can it be?
3:14
Is this really reproducible?
3:18
And last of all, AI analysis for to analyse immuno biomarkers.
3:24
That's all very nice, but it's extremely complicated to set up in a hospital environment and you will always have, you know, you need big budgets.
3:33
This is only for the lucky few and I hope to prove this wrong and just to go back to the reason why we have precision medicine in the first place.
3:45
Many people think about Herceptin or Tarceva.
3:50
Actually the first and most important, I think diagnostic biomarker is linked to tamoxifen and it's a very interesting history.
3:57
So Dora Richardson was a chemist working in Manchester at ICI Chemicals.
4:03
At that time, she was working on a new class of oestrogen inhibitors working, presuming that this could be used as a contraceptive.
4:12
But it turned out that this actually worked in breast cancer, but only sometimes.
4:19
And understanding the mechanism that the oestrogen receptor needs to be bound by this inhibiting molecule, and maybe some cancers do not express an oestrogen receptor, led to the first oestrogen receptor test, which was a radial ligand binding test.
4:37
And it quickly converted into an immunohistochemistry test.
4:42
Why?
4:42
Because that's much more accessible.
4:45
But it took an amazing 35 years to actually prove in a meta-analysis that it's actually a valuable therapy in combination with an oestrogen receptor test.
4:58
Look, just look at the dates of these publications.
5:00
They're all late 80s, early 90s.
5:04
And this is really the dawn of immunohistochemistry as a technique.
5:10
Remember in the beginning this only worked on fresh tissue, and it only was when people discovered antigen retrieval that this could be used on formalin fixed paraffin embedded tissue and quickly improvement of mouse monoclonal antibodies.
5:28
The advent of diagnostic kits, automation of staining. And today image analysis led to the fact that this is really especially compared to molecular techniques a routine test done in almost all pathology labs with a two to three day turnaround time.
5:47
It's reimbursed so there is no issues like this is a $5 million test can only be done in two places in the world.
5:55
This is really commonly accessible. And if you look at, this is a lots of information on this slide, but this is a screenshot of the latest NordiQC quality control round for HER2.
6:10
There were 400 participants in this round, and you can see the over time the result of labs obtained for different diagnostic tests.
6:19
You can see the top performers are very stable and then you have the green line on the bottom that's very variable, but you should know that this represents just a few users out of those 400.
6:32
So the majority of labs are using very reliable, very robust kits and to stay in the movie team oestrogen receptor, HER2, ALK, PD-L1, Claudin18 and there's a number of other ones.
6:50
The bulk of precision medicine drugs therapies are prescribed based on the result of immunohistochemistry test.
6:57
That is really a majority, I would estimate this at probably at least 80% of precision medicine therapies.
7:07
So this leads us to the following question, especially PD-L1. I was closely involved in the launch of PD-L1 as a diagnostic test.
7:17
How do you train all those pathologists on this complex algorithm?
7:21
And is it really a reliable reproducible technique even with, you know, just a few days training?
7:30
So one of the difficulties to do training is that you need slides to be present all over the world.
7:36
So the ideal format to do it digital pathology.
7:39
So this is basically what Pathomation does for these kind of studies.
7:43
We set up a platform, be populated with well-defined slides, these are consensus scores and then put in a format where people can either go for self-learning or can be used in a classroom setting, test themselves afterwards and generate a certificate that they are qualified to do this.
8:07
And this is one example of a use case for PD-L1.
8:13
This was done with one of our partners, the these were trainings done over 4 continents.
8:23
So in Asia, Europe, South America and North America.
8:28
It was a preliminary setting.
8:29
So the study size is not that big.
8:31
It's 70 people, but they were trained to score across the 1%, 25% and 50% and tumour proportion score.
8:42
And it is with an overall reproducibility of within the cap limitation, which was 85-95%.
8:52
And remember, this was very early days.
8:55
People have not seen PD-L1 stains on lung cancer.
8:58
It looked somebody said this looks like the dog's dinner, which is an expression in English I learned that day.
9:07
Apparently it means very confusing, very dirty.
9:13
But today we can do the same thing, but now with AI assistance.
9:16
So what you will see here, it's a video.
9:20
I hope it plays.
9:21
So this is a learning environment where people can navigate through these slides, and they have the information about what is the exact result next to that.
9:31
And they can actually call upon an AI analysis algorithm for HER2 in this particular case, which will show you in detail, you know, how many slides we're staying for which specific intensity, and this overlay indicates where those cells are.
9:56
So this could be, for example, useful in the new setting for HER2 ultra-low.
10:01
We'll come back to that in a moment.
10:04
Other way around.
10:05
Imagine you are in an exam situation and a test situation and afterwards you get your results.
10:12
You know, you're always curious if you get an answer wrong, where did I go wrong?
10:17
Please explain what I did.
10:20
So this is basically showing how people run through a self-assessment trajectory.
10:29
You basically indicate, I think this is a two plus, I think this is a three plus.
10:35
When you complete the series, let's give it a second so you can download a report which shows you where you were correct or wrong.
10:51
And this can be populated with more explanations.
10:54
But I think the logical thing to do is look at the AI generated result and use this for, you know, checking yourself, you know, why did I give a different result?
11:06
This obviously can also be done with a pathologist helping you and guiding you, but we know how busy pathologists are, so this could be a nice adjunct to those specific situations.
11:21
OK so as I said, new diagnostic challenges that lie ahead.
11:28
The first one probably is as you all well know very well, there is a new anti-HER2 drug and a cytotoxic compound is linked to it.
11:40
Enhertu is I think the commercial name.
11:44
And it will actually also work on lower expressing tumours, and it will divide the previous 0 category and two new categories.
11:55
So absolute zeros and the ultra-low which also respond to therapy.
12:01
So this requires some retraining.
12:04
So it's a very interesting situation to train pathologists again.
12:09
But what I have heard from, and this was a poster last year at ASCO, the current HER2 assay is in combination with the scoring algorithm, probably not very precise enough.
12:23
And Doctor Navarro used this analogy.
12:27
We are trying to weigh mouse on a scale that was designed to weigh an elephant.
12:32
So this is a potential use case for AI analysis and routine.
12:36
But probably the real one that will change everything will be this algorithm.
12:43
So it's the, it's specifically for TROP2a where this is something the human eye kind of do.
12:50
You're looking at optical density and you're comparing the optical density and the cytoplasm over the sum of the membrane over the sum of the optical density and the cytoplasm and the membrane.
13:04
And this will give you a ratio.
13:13
Here, the normalised membrane ratio and if the majority of tumour cells is has this ratio below a certain value, it would be a reason to prescribe this specific therapy.
13:31
But labs need an AI menu.
13:35
You can think about all the different AI vendors that are available on the market.
13:40
Maybe you have one in mind for prostate, another one for your HER2 algorithm and the QCS algorithm is yet another provider.
13:50
So how can you manage that too?
13:51
You need to go into a relationship with four to five different providers, or you know, in the future, as we have seen, unfortunately in the last months, will there be bankruptcy?
14:04
Will there be consolidations?
14:06
There's two models and remember, I'm from Belgium, so ABInBev.
14:11
It's now the biggest regrouping of Brewers.
14:14
They basically keep all these brands and put them together.
14:19
And the other model I have that is a potential and for me this is the winner is the App Store model where you can go to a certain website and use the different AI algorithms.
14:30
So we strongly believe in the last model, and we have basically developed one of our current SaaS platform to include this capacity.
14:42
So instead of trying to integrate the AI algorithms in your hospital directly in the clinical setting, you can do something else.
14:50
You can upload your image to this SaaS platform and choose from the menu the specific AI tool you would like to use on this.
15:00
I'm not saying this is for high volume settings, but it's ideal for to get started for intermittent use or for those special cases where you just feel insecure.
15:11
By the way, at the same time, the platform also allows to consult a colleague and ask them for their opinion on the specific case.
15:20
So conclusions.
15:22
I think it's very clear.
15:24
You cannot understate the importance of immunohistochemistry.
15:27
It's very reliable, it's broadly accessible.
15:31
Pathologists are used to these techniques.
15:33
They can with proper training do this very reproducibly across different sites, you know, across in time.
15:43
And I think the AI image analysis is really the next thing to do, especially the QCS algorithm will be very important.
15:51
But I think this really depends on you on training.
15:54
So AI is a new tool, something you would like to use, something you need to get used to, you know, remember what I tried to do with the image, the movie poster?
16:06
But also it can be a bit unpredictable.
16:10
One of the things that struck me also is this very strange analogy in the background.
16:17
Somehow, if you remember the movie Terminator, you had this striking resemblance between the eyes of the Terminator and the character the AI generated.
16:28
And that was all I had to say for today.
16:31
Thank you.