Thank you for this invitation to join Lunafore’s spatial biology week. My name is David Mason from Visiopharm, and I’m going to be talking about an automated image analysis pipeline for highplex sequential immunofluorescent images.

Now high and ultraplex data are becoming much more common, but image handling and analysis often prevent some very specific challenges.

These can be things such as a workflow that requires multiple platforms.

This is less than ideal because whilst each platform may do its job quite well, this will involve exporting and importing data between the steps, really inhibiting an end to end workflow.

We also hear that cell detection can often be unsatisfactory.

This is also if it is using a traditional or intensity based approach to detect the individual phenotypic units.

These are complex datasets and can be very time consuming to perform image analysis and QC.

And lastly, something we hear quite often is that users can lack confidence in their results, not being able to explore and QC their results in the way that they might like.

Throughout the purposes of the talk today, I’m gonna be using LUNA4 datasets, including this twenty two plex plus DAPI tissue microarray that I’ll use for all of the examples in the rest of the talk.

Now hopefully by the end of the talk, you can appreciate how Visiopharm can be used to address all of these and more of the challenges in analyzing complex datasets.

But first, a brief introduction to Visiopharm as a company.

Visiopharm is a market leader in computational pathology.

We’ve been around almost as long as digital pathology has existed. We’re over twenty years old and have all those years of tissue experience.

This has given us strong ties into the scientific community and appreciation for what is really required by the field.

We’ve also used this time to develop cutting edge technology and have many patents to our technologies and using the AI techniques that we have to create a best in class platform.

Another key pillar of the software is that we develop flexible solutions.

This means that you’re not locked in to a single solution and have to come back to us for more. We give you the toolbox, and you have an infinitely configurable set of tools to solve almost any research question.

Lastly, something we believe very strongly in is our expert support team. They’ll help you not only to learn the software, which buttons to click, but in application support and training as well to solve your bespoke problems with the Visiopharm platform.

Despite being around for twenty years, we’ve kept a very lean team with just over a hundred people worldwide. And as you can see from the map, we actually have global presences and service the entire world.

So how can Visiopharm help you solve some of the problems in today’s digital pathology field around highplex and spatially relevant data?

I’m going to touch on three main topics during this talk. The first is pretrained knowledge.

The second is the concept of accessible deep learning, and the third comes back to this previous point, which is flexibility in the tools and how you use them.

So let’s start with the concept of pretrained knowledge.

Users of Visiopharm have access to over a hundred and forty pretrained, prebuilt apps available in our app center. You can see the URL on your screen, and you can go and explore all of the different tissues, stains, and applications that we have examples for.

These apps can be used off the shelf on your own data, or, alternatively, you can use them as a training reference, or oftentimes, you can take the existing app that’s been built and then refine it with your own data.

Three apps that I’d like to particularly highlight, and one of which I’ll use in the purposes of today’s talk, are the nuclear detection apps.

The three main modalities that we see in Visiopharm are fluorescent imaging, brightfield imaging, and imaging mass cytometry. And for each one of these, we have a deep learning app prebuilt to detect individual cells in these different staining modalities.

This significantly reduces the amount of time it takes for you to develop your own workflows and your own solutions because these have been pre pretrained across thousands of images and so are very good at detecting variance across these datasets.

More on that later.

I’d like to take an example from this LUNA four dataset we have here using these pretrained knowledge tools in order to do a nuclear detection.

The first step, of course, is to pick your input channel, and for the most part, we would be using a DAPI channel. We can, of course, also use other stains, histone two b or CITOX.

And then we simply run the pretrained app over these datasets to detect individual cells or phenotypic units.

I would just highlight that using deep learning to do our cell detection gives us really robust detection of these individual cells as deep learning excels at handling variance, not only in intensity, but also in size and shape.

And this is perfect for real world tissue where you frequently have large tumor cells, small immune cells, stromal cells, fibroblasts, and so on and so forth. Deep learning is able to capture the variance across all of these different cell types and then present it back to the user.

So once we have ourselves, how do we go about phenotyping these individual datasets?

Again, I’m using this Lunaford TMA to as an example, but it’s really a very simple process. I’ve highlighted it here in four main steps.

Number one, as we just discovered, is using the off the shelf tools to detect our individual cells.

The second step is to then pick from your individual channels and decide what you would like to phenotype.

In many cases, for instance, we want to exclude things like the DAPI marker or perhaps structural markers like, cytokeratin or CD thirty one.

We can simply select which channels we’d like to use, for instance, in this case, a simple t cell panel, and then hit the train button within Visiopharm.

This will then learn using a machine learning approach and then provide you not only with a list of singlets that have been detected, but also the more complex phenotypes. You can see here we’re detecting FOXP three positive CD four cells and FOXP three positive CD eight cells, as well as, non CD four CD eight FOXP three and non FOXP three positive cells.

Using this phenotyping, we can then run this back on the original dataset to quality control and review right there on the image.

Once we’re happy with these results, we can, of course, upscale and batch process an entire dataset. The app no longer needs training because it’s already learned what each of these different phenotypes looks like.

I’ll just take a moment here to reiterate that what Visiopharm aims to do is to provide an end to end solution that allows you to upscale your work and process large datasets.

This is really valuable from a reproducibility perspective because if you can change inputs or change your phenotyping parameters and then reprocess an entire dataset very easily, it really helps with the QC cycle.

So we’ve phenotyped our our cells and our data. What can we do with that information now? Well, Visiopharm allows you a real, flexibility about outputs as well. So I’ve listed off some of the outputs you can produce in Visiopharm, of course, things like counts, if you wanted to count all of the CD eight cells, if you wanted to calculate percentages or fractions. What fraction of this total CD eight population is FOXP three positive?

Because we can also measure the area of our tissues, all of these numbers can be converted into densities, allowing the number of cell pop of cells of a particular population per millimeter squared, for instance.

When we’re detecting cells, we also have access to a whole raft of morphological, outputs as well. Simple things like area, but also shape, concavity, circularity, solidity, many other things that describe the shape of the cells.

We also have a lot of spatial outputs. And in fact, almost anything you’d want to measure in Visiopharm, you can, including just simply exporting the coordinates of each individual cell, measuring distances between phenotypes, for instance, how close are the macrophages to the t cells, but also the distance to boundaries. And I’ll come back to this a little bit later. This is great for looking at, tumor infiltration or invasion from a blood vessel or so on and so forth.

Lastly, you always have access to the raw intensity values. We can also look at things like variance or texture features if you’re interested in large scale data mining or things like cell painting.

And the other important thing to mention is that any of these numerical results, which can be exported at the cell level, can also be accessed by exporting as text files or the overlays as masks as either XML vector data or TIFF files. So you never get locked into the platform. If you wanted to go off and do your own clustering analysis or your own spatial analysis or visualization, you always have access to the data from Visiopharm.

So this really deals with the first element that I wanted to talk about. But I wanted to add an extra layer onto this by talking about Visiopharm’s accessible deep learning tools.

In Visiopharm, if you can’t find an app that does what you want to do from our app center, you can always create your own using a using a train by example deep learning.

This works on any image type, including fluorescence, imaging mass cytometry, basically, anything that can be arrayed as a spatial image.

This also leverages the most powerful technology in AI, which is the deep learning approaches.

All you need to do is use your biological knowledge to create apps for object detection or region segmentation. And I’ve given a couple of examples on the left here. For instance, kidney glomeruli, layers in lung, fibrosis, tumor, or even detecting necrotic or artifactual areas in your image.

So if we go back to our LUNA4 dataset, we’ve started talking about cell detection, but we can do so much more with these images. We have so much information in all of the twenty or thirty channels that are available.

So how would we go about, for instance, splitting this image into areas of stroma and areas of tumor?

As I mentioned, it’s a train by example approach, and all you need do is to give examples to the software of what tumor looks like and what stroma looks like in your particular images.

You can create these annotations very easily in the software and then do this across, multiple individual images to show different presentations of tumor.

Using that same train button as we saw before, it’s very easy to create, a new app. You’ll get live feedback on how well it’s performing, and then you’ve got your own app to perform tumor detection.

It really is this simple.

The really powerful thing here comes when we’ve created this app to do tumor detection is that we can then daisy chain in other apps. So going back to our previous workflow, we can do our tumor detection and then perform cell detection, opting to analyze only the tumor, only the stroma, or both as in this case.

And then we can perform our unbiased phenotyping using this same boundary information.

This is really, really useful at this point because what this then allows us to do is to create some of those basic outputs that we were talking about before in the context of spatial. So if we wanted to count the number of T cells, for instance, we can now count the T cells in the stroma and in the tumor separately.

We can also perform these density analyses or look at how many tumor segments there are in our individual images.

And as I alluded to before, we can start to look at the relationship between our phenotype cells and those boundaries, looking for instance at tumor infiltrating lymphocytes or the distance cells have traveled from blood vessels for instance.

In the last part of the talk, I’m gonna take what we’ve learned so far about pretrained knowledge and accessible deep learning using our workflow of image analysis, and just highlight how these tools can be used to answer some really interesting and flexible questions around spatial biology.

I’m going to start looking at, a spatial example whereby instead of simply looking at the tumor areas, what we’re actually going to do is to take it one step further.

We wanna initially detect the tumor, and I’ve shown you how we can do that with train by example simplicity.

Following this, we can use tools in Visiopharm to actually dilate our tumor boundaries.

What this is doing is identifying a particular area of interest within, say, twenty to fifty microns of our tumor boundary.

This is more like an invasive margin or an area of the tissue that’s responding to the presence of these solid tumors. Of course, once we determine these areas, we can follow exactly the same steps as seen before by phenotyping the cells only within a particular region.

This allows you to ask some really interesting spatial questions, for instance, as to what’s going on proximal to the tumor areas.

Of course, we can choose to dilate out from the tumor, we can dilate in from the tumor, or select regions that are the furthest away from the tumor. It really is a flexible approach to answer your questions.

Another approach would be to look at macrophages, for instance. So in this particular case, we’re leading with the the the known knowledge that m two macrophages, a particular type of macrophages, have been known to induce replication or tumor growth in some cases or tissue remodeling.

And so how can we use Visiopharm to answer those sorts of questions?

Again, the easy step by step process would be to come in and detect individual cells in the tissue.

We can then classify the cells as m two macrophages. In this case, we’re simply looking for CD sixty eight and CD one sixty three positivity.

You can see these are these orange cells in the image to the left.

We can then ask for all of the individual cells that are proximal to our m two macrophages. In this case, I’ve chosen less than five microns. If you wanted a slightly wider boundary, you’re free to adjust this.

Once we’ve detected those proximal cells, for instance, we can then classify them based on their Ki sixty seven status.

And what you can see in the graph, plotted here is that for this particular core, indeed, when compared to the global proliferation index of about ten percent of cells, the m two proximal proliferation index is almost double that.

Using this concept of building up an analysis and then running it on a single image, we can, of course, extend this to the larger datasets as I mentioned before.

We can batch process all of the calls on a TMA or all of the images in a larger dataset to get results for tens, hundreds, or thousands of individual slides that can then be exported and then processed statistically.

The last thing I’d like to mention, when talking about flexible tools and one of the last points that I introduced right at the beginning of the the talk today is actually around, data exploration and quality control.

A lot of the analyses that I’ve been demonstrating today, will often end up with datasets that really warrant exploring and QCing.

And in Visiopharm, we have a data exploration tool based on tSNE plotting or box plotting or scatter plotting that really allows you to explore your data.

What you can do is to take almost any input from the calculations in Visiopharm, areas, cell size, shape, morphology, spatial parameters, all of these things I’ve been talking about, including intensity, which is one of the most common, and then have those, put into a tSNE plot to reduce their dimensionality.

The aim here is to really identify what sort of parameters will cluster cells together.

You can see on the right that any of those input parameters such as intensity or shape or size, you can heat map for each of those parameters. And you can see there’s reasonable separation between the three, different parameters that we’re we’re, plotting on here.

This is a really dynamic plot. So this also allows you to recolor the individual, groups, but also to identify particular cells. So here, we’ve selected some cells just by clicking into the tSNE plot, and it highlights on the image where those cells are. This is great for identifying things that cluster around boundaries or in necrotic areas, for instance.

Lastly, this, interaction works both ways. If we click on an individual cell in the image, it will then jump to where that is in the tSNE plant. So we can highlight a particular cell and make sure it’s aligning with what we think is the correct, phenotypic allocation.

So just to review quickly, I started on these three main pillars of Visiopharm.

Pre trained knowledge, using these off the shelf apps really saves development time. You don’t have to do this yourself when we’ve done it for you already, and that gets you to your results faster.

Accessible deep learning only requires the biological knowledge to train a bespoke app. You can teach someone to use these tools in about five minutes. At that point, they can then go off, create annotations, and train their own deep learning apps without the need for any scripting or coding or complex software installs.

And lastly, one of the main pillars of the Visiopharm software is this concept of flexibility.

We wanna give you the tools to ask the questions you want answered, not the ones that we think you should be answering.

You can also, as I demonstrated in the latter half of the talk, customize the workflow, ask some really interesting questions around proximity, invasive margins, different regions of your tissue, and then get access to all of those results.

Just to summarize this whole talk, Visiopharm has extensive pretrained knowledge to reduce time to results in your studies.

We’ve got easy to use train by example and a multiplex classifier, which automates classification and phenotyping.

As I’ve mentioned before, this flexible toolbox allows you to answer almost any questions, especially around spatial biology and interesting biological phenomenon.

And lastly, as I mentioned right at the beginning of the talk, we have an expert support and training team, so you never get left alone. We’ve always available to help you with training in the platform, but also application level support.

I really appreciate your time today. I look forward to the question and answer period that’s coming up. If you have any questions that don’t get dealt with in the q and a period, please feel free to reach out, to the email that you see on your screen now.