0:00
Good afternoon, everyone.
0:01
So I'd just like to start by giving a brief introduction to Oracle Bio for those of you who may not know us.
0:06
So Oracle Bio provides quantitative digital pathology services to support pharma and biotech R&D.
0:14
The company is coming up to 15 years old.
0:17
We have a global client base covering Europe, Asia and the US and we work right along the R&D pipeline.
0:24
So we support studies from early discovery, preclinical right through to late stage clinical trials.
0:31
Now it's an exciting time to be working in digital pathology and especially in its application within pharma R&D.
0:40
There's been a huge amount of innovation over the past few years and this has led to some significant improvements and things like data quality, workflow efficiencies, but also the degeneration of novel insights to help drive R&D forward.
0:57
However, it can be challenging to state keep the pace with all this constant innovation in the field and also to effectively apply it within an R&D setting.
1:09
So in today's presentation, I would like to give some examples of how innovation and quantitative digital pathology is positively impacting image analysis and data generation workflows at Oracle Bio to support our services.
1:23
I'll also talk about some of the challenges we come across in implementing utilising, keeping pace with quantitative digital pathology.
1:31
And I'll finish by giving my perspective on how we believe a more integrated working relationship between pharma and CROs in the digital pathology space can deliver valuable synergies to optimise the impact on R&D.
1:45
So this is a standard image analysis workflow and it's one that we utilise at Oracle Bio.
1:52
And in the coming slides, I'd like to talk through each of these areas and to just give examples of how innovation is impacting on these in our organisation.
2:02
So starting with image transfer metadata QC, standardly images would be sent to us from a starting location.
2:10
So for Oracle Bio, that would either be a pharma company or a Histology CRO sending us the images and they would be transferred into our AWS into a file in our AWS and then those images would be loaded into our image management system for starting that process with our pathologists and our image analysts.
2:28
However, there can be some issues that arise.
2:31
It could be an issue with loading or opening the image in our system.
2:35
Could be the wrong file format, it could be an incorrect scan magnification, or it could be for Multiplex IF images you have the wrong channel order channel names and this just delays the process of working the images through the analysis.
2:49
But it also uses up valuable time of our scientists and our pathologists.
2:53
So at Oracle Bio, we've added in a few pre steps before loading images into our IMS.
3:00
First of all, when images come into a folder in our AWS, there's an automated virus scan runs.
3:06
And that's just to ensure obviously that anything coming from outside or organisation into our digital pathology environment is clean of viruses.
3:14
We can also perform checksums.
3:17
So checksums is a programme that takes a digital fingerprint of all the information in a file of its bits sizes and creates a value.
3:26
And that can be generated at the starting location and also at the end location.
3:31
Those values need to be identical because if they're not, it means there's been some corruption in the transfer of that image.
3:38
And then thirdly, we also do an automated file metadata QC.
3:42
So here at Oracle Bio, our R&D team have generated a web interface underpinned by Python, which means our scientists can now go to that folder very quickly with the images and open it in this web interface and it lists all the metadata for the images in that folder.
3:58
So we've things from image name, file format, magnification, bit depth, channel numbers, channel names and order.
4:05
And it also identifies any outliers in the columns.
4:08
So for example, in the file format, you can see there's one file there that's different to the others.
4:12
So it's very quickly identifies any issues in the metadata.
4:16
So when we have any of any issues in these steps, these pre steps, they can be identified quickly and fed back to the starting location to be addressed.
4:25
And only images that are of high-end quality and passes can then be moving into our image management system for work by our image analysts and pathologists.
4:36
Just moving on to image QC and annotations against traditionally we would have loaded images, our image analysts would have loaded images into our image management system or software and we would have manually annotated the tissue present on images, but also the artefacts as well.
4:52
So they're negatively annotated to exclude them from downstream analysis.
4:56
And you can see some examples here of tissue falls and some histological deposits, tissue lifting and so forth.
5:03
And you can imagine that if there's a lot of artefacts on the section, this takes a long time to annotate.
5:11
Obviously if sections are bigger, that also takes longer.
5:14
And if you've got hundreds of images in a study, this is really quite a time intense labour intensive process.
5:20
So more again, recently we've started to use AI driven automated tissue and artefact detection module.
5:27
This is the Halo slide QC version three we use indica labs Halos, one of our platforms.
5:34
And what this does, it automatically runs over the images now in a folder and very quickly creates a mask for viable tissue and artefact.
5:43
And those artefacts can then be excluded from any downstream analysis.
5:46
And as you can see in the examples here, it's doing a good job and picking up things like those prominent artefacts on folds and deposits.
5:55
But also in the top right image there you can see there's also one of the tiles of the image is out of focus.
6:04
So it's also detected things like that which can be difficult to do by eye.
6:08
Overall, this is version 3.
6:10
It's been trained more than the other versions from Halo.
6:14
It's getting good generalizability, but it might not necessarily work on all images from your specific studies or our specific studies, but it can be further trained with images from your own organisation just to build in that extra generalizability.
6:29
Moving on to algorithm development.
6:31
This is probably one of the biggest areas where innovation has had an impact over the last number of years, especially with the integration of AI.
6:41
And I'd like to talk about how we use this in an R&D setting.
6:44
So at Oracle Bio, there's normally 4 steps with respect to algorithm development for R&D.
6:50
First of all, there's the tissue and artefact detection, which I've touched upon in the previous slides just to say here that can be applied to whole sections, but it can also be applied to a pathologist annotated area on those images.
7:03
The next step is tissue classification.
7:05
Here that we're taking that detected tissue and further classifying it into particular regions of interest.
7:11
What?
7:11
So for something like a cancer sample, it could be tumour stroll now, but also necrosis and glass.
7:17
The next stage is N cell detection.
7:19
Important here is to be able to detect cells in different sort of staining intensities and different tissue morphologies, but not just to detect them, but also to accurately segment those cells in situations where there's high densities of cells or cells that are overlapping and so forth.
7:35
And then finally by adding in the thresholds we can start to, we can start to generate positive, negative cell populations and then further classify cells into low, medium and high intensity for those that are positive.
7:49
Now at Oracle Bio, we use commercially available software.
7:53
We use Visiopharm and indica labs Halo as two of our main platforms.
7:57
A number of pharma companies also use these, a number of our clients use them.
8:01
And as you will see from their website and their offerings, there's a lot of AI algorithms coming through from these organisations to really support R&D and research in a pharmaceutical environment.
8:12
And that's great to see.
8:15
In general, these algorithms cover things like tissue classification, cell detection and they will generally work well on a number of different examples of images based on what they were, what those particular algorithms were trained on.
8:29
However, with R&D, we are dealing with quite a diverse selection of studies that we see on a day-to-day basis from our clients, different tissue samples or even for oncology, it might be different cancer types in that study, different staining modalities, immunohistochemistry and such a hybridization, multiplexing and different data required from those image really makes it quite bespoke a lot of these studies.
8:53
So what we tend to find at Oracle Bio is that all those algorithms are good starting points.
8:57
We then need to further refine them for those specific studies that we receive from our clients.
9:03
So an example of that would be something like this.
9:05
And I appreciate these are TMA cores, but it's just to give you the example that these are all gastric TMA.
9:12
And as you move from left to right across here, you can definitely see that there's differences in the tumour architecture, tumour stroma content.
9:20
There's also differences in the amount of tumour staining for the marker and brown.
9:26
And there's also some evidence of areas of necrosis and background staining.
9:30
So that when we get a study like this in an R&D environment, what's important to do is to select images that represent the morphological diversity and the staining heterogeneity of your sample set to train.
9:44
We will then, along with our pathologists, go in and annotate areas that represent the different classes that we want to detect.
9:51
And then through an iterative process of training and review, we will get to the point where we are developing a nice tissue classifier algorithm that is working across these kind of broad spectrums of morphology and staining heterogeneity.
10:07
It's not always the case that we can develop this all in one algorithm.
10:10
And it's really important to recognise that there may be situations in a study where the cancers are so diverse that it's not efficient or effective to try and build them all into one algorithm.
10:22
So again, just as an example, we've got small cell lung cancer here, Melanoma and glioblastoma.
10:27
And as you go from left to right, you can see the tumour stroma architecture differs, but also the staining of the target differs as well.
10:36
So what's important again, from our perspective as a CRO work in this space, we need to come up with a strategy.
10:42
We will look at all the samples in the study.
10:45
We will then see how many algorithms we feel are suitable to generate.
10:50
It may be that 80-90% of the samples in the study are captured by one algorithm, but there may be 10 or 20% that aren't.
10:57
And what's the important distinction here is that these studies are data focused.
11:02
They're not algorithm tool focused studies.
11:06
So whether we need 3 algorithms to get high quality data, that's OK in an R&D environment and amendments are allowed.
11:14
Whereas in healthcare, the focus is very much in developing 1 algorithm that's going to work across all.
11:18
It's definitely much more efficient and sometimes more accurate for us to take this approach in an R&D environment.
11:26
Just moving on to sort of cell detection, again, I would say that vendors that we work with and I'm sure other vendors have put a lot of effort into continually improving nuclear detection and segmentation in their algorithms by broadening the training of different tissues and different staining modalities.
11:43
And in general, they're working pretty well out-of-the-box.
11:46
But again, within that R&D environment of bespoke studies, there will be a number of times we will come across studies where we have to do some further training of the algorithm.
11:57
And you can see down in this bottom corner here, I've just annotated some nuclei.
12:01
So we would test the algorithm and check and see how well it's working and if we need to in areas where it's not working so well.
12:08
An example of that sort of how it works is in the middle panel for nuclear detection.
12:12
We will further train that algorithm just to refine and to fine tune that overlay.
12:19
Once we have developed something that's optimised, we can then move with confidence into adding the thresholds and the cell compartments they need to generate that high quality data.
12:28
And then finally, from an algorithm perspective, I did want to also touch upon the hot topic of cell compartment analysis.
12:35
It's something that we get a lot of requests of now and this involves and there's a need here to generate membrane compartments as well as nuclear and cytoplasm.
12:45
And if we go back to the time before AI was being used in anger, it would have been very difficult to generate a membrane analysis without AI.
12:55
We have these scenarios where membranes aren't fully connected.
12:58
We have different variant levels of staining across the section for membrane staining.
13:03
And we can also scenarios where there's cytoplasm as well as membrane staining.
13:06
So certainly AI has made a huge impact on how we detect membranes and these algorithms are getting better all the time.
13:13
We will, again, based on the specificity of the study, the tissues, the heterogeneity is staining, sometimes have to further retrain the algorithms to make sure that overlay is accurate for that particular study set.
13:24
But when combined with the nuclear AI, you can see we're now generating 3 compartments per cells.
13:30
We're able to threshold and build data for membrane, cytoplasm and even nucleus as independent parts of the same cell.
13:37
This allows things like membrane to cytoplasm ratios.
13:40
And that's a really big area in things like ADC and companion diagnostic studies that are going on in cancer.
13:50
OK, just moving on to analysis processing.
13:53
I've just got one slide on this, but Oracle Bio has gone through the journey of IT over the last 15 years.
14:03
We started on desktops, moved to local servers.
14:05
And about 6-7 years ago, we went into the cloud, went into cloud infrastructure, and it has been transformational about how we help we perform analysis and deliver data back to our clients.
14:18
We now house all our softwares that we use up in the cloud very close proximity to the images that are also in the cloud.
14:25
And this means that we're really then able to quite effectively leverage all the power and value of AWS and the GPUs and CPUs that sit there to be able to perform scalable analysis and perform batch processing of images in parallel.
14:39
It does mean we're able to generate data from tens to hundreds of images within a day and deliver that data back to our clients at Oracle by we got a lot of help from our sister company Ciento in developing this cloud infrastructure.
14:54
If anyone has any interest, please do come to us at our booth and we can pass on more info.
15:00
And then finally data management and QC again, equally still an equally important part of the process, but when data is generated in analysis software, it's normally exported as a CSV file.
15:11
That CSV file in our hands is normally rendered down into a more digestible Excel file which would be passed back to our clients.
15:20
But through that process, there can be certain errors in developing that Excel file, like copy, paste, formula, formatting, data, QC.
15:28
So what we've done at Oracle Bio as well is again, with our R&D group, we've been able to develop a web interface where our scientists can now import the CSV files directly into this web interface.
15:38
And you can see the columns along the bottom, in the top right, you can see all the column names and they can be edited, they can be moved around.
15:47
They can be deleted in this interface.
15:49
And we can also on the left, top left, we can add extra columns.
15:52
So if we wanted to add columns with a formula, for example, numbers of cells per millimetre square root of a region of interest, they can all be added in.
16:03
And then Python underpinning this will do all the data rendering in QC for us so that we export that data out as one file.
16:11
And it's just reducing manual errors and making it a more efficient process for us.
16:17
So overall innovation is having quite an impact on Oracle Bios workflow, image analysis workflow from the early steps of that pre cheques on the images before they go into our image management system, The ability now to bring in automated slide QC, the fact that now the IMS systems we use Halo Link is opening up to have Visiofarm and hopefully more third party algorithms coming up.
16:41
That's a great move because it allows us to use our image analysis toolbox more effectively with our image management system and then right through to having Python supported data curation and QC to deliver the data.
16:54
This isn't a finished product.
16:55
It's a live evolving environment and no doubt it will be refined and added to as innovation continues to have an impact on our workflows in Oracle Bio.
17:06
So how can we as a company help our pharma clients in this highly innovative time?
17:13
Well, first and foremost, we've always offered fee for service model, which is where we receive individual projects with a defined remit from our clients.
17:23
And we will perform that work with all the innovation and efficiencies that I've just demonstrated inside our organisation to deliver that data in the best way possible.
17:34
However, in the last couple of years, we've also introduced our FTE model.
17:38
Now here clients can now access Oracle bio staff or expertise on a time basis, 36 or 12 months.
17:46
And that can be for 20 days a month or for half an FTE ten days a month.
17:51
And what this is, it's not access to one individual in their organisation, but it's an access to the amalgam of all the capabilities in our organisation.
17:58
So if you need a Halo expert, you can get access or if you quickly need to change to a Visiopharm expert, a Python scripter, one of our clinical pathologist or web, this web interface designers, you have access to that.
18:10
In this model, the work can be done in Oracle Biosystems, but equally for a number of our clients, we're actually remotely deploying our scientists into their digital pathology environment.
18:21
So all the images that we work on, all the algorithms we generate, all the data we create stays in your organisation.
18:28
It's flexible, you can change priorities based on your needs very quickly.
18:34
With the FTE, there's no paperwork, there's no contract amendments.
18:38
Plus you can also get our FTE to work on full studies.
18:42
Or you might just want them to work on QC and develop some algorithms for you, or to build some of the workflows that we've developed internally to support your internal infrastructure.
18:54
And it's a shared learning environment.
18:55
We're working directly with the scientists in your organisation to problem solve, but also to share best practises and also to give sort of updates on best ways to use new software or new updates of software versions.
19:13
And then finally, clients can come to us directly for clinical trial studies and many do, but also if you are working with IQVIA, we now have a partnership with IQVIA Laboratory Services where clients can have their staining and imaging done through IQVIA laboratories, but the image analysis can then be performed through Oracle Bio and that's all through the IQVIA contract.
19:38
And through that we can develop the algorithms and run them on those studies.
19:41
Or actually you can just send us your algorithms from your own internal activities and we will run those on our clinical trials on your images.
19:50
And that's for both exploratory and GCP studies.
19:55
So in summary, I hope I've demonstrated the QDP innovation is positively impacting image analysis and data workflows at Oracle Bio.
20:04
And we also see that across our wider client pharma community to help drive novel insights and efficiencies in R&D.
20:12
There are a lot of challenges and I'm sure there always will be an implementing, utilising and keeping pace with quantitative digital pathology in an R&D setting.
20:20
However, our experience is telling us, especially through our FTE model that a more integrated working relationship between pharma and CROs can deliver valuable synergies to optimise the impact on R&D.
20:34
So with that, I would just like to thank you for your time.
20:37
I hope you found that interesting as well.
20:39
If you want to come and speak to us about anything, we're at booth 60 and I'm happy to take any questions as well.