Hello everybody and welcome to today’s webinar titled Improving Diagnostic Workflows with AI, Lymph Node Metastasis Detection in Breast and Colorectal Cancer.

I will be your moderator today and my name is Bettina Winkler.

Today’s speakers are Stine Hader and Dzenita Omanovic.

Stine Harder is our Head of Clinical Product Management and she is crucial to product development here at Visiopharm. Dzenita Omanovic is one of our Image Analysis Specialists and she is highly skilled in CE-IVD apps and data sourcing and management.

She specializes in external collaborations and projects with the clinical CEIVD or future FDA scope.

So with that I will hand it over to our speakers.

Thank you Bettina.

Lymph node status is a major factor for clinical outcome of those two cancer types, and detection of lymph node metastasis is therefore critical for treatment decisions.

Lymph node assessment is is usually performed by an expert pathologist using one or often multiple H and E stained tissue sections per lymph node.

Given that several lymph nodes are often assessed per patient, the workload quickly increases for the pathologist.

The current workflow is tedious and time consuming, which might even lead to an increased risk of an accurate diagnosis.

This is how the current workflow for evaluating lymph node status looks today.

Lymph nodes are excised, sectioned, fixed, and stained.

The full tray of glass slides are then provided to the pathologist that screens all slides for metastasis and reports the number of positive lymph nodes.

For the pathologist, this workflow means spending a long time in front of the microscope and manually keeping track of slides and slide results.

The manual assessment thereby becomes time consuming and costly.

Visiopharm’s Metasys detection app for whole slide images can easily be integrated into your workflow, Designed to optimize your review process, so you can make the decisions with confidence.

After the lymph nodes are processed, Visiopharm’s software identifies and highlights metastatic regions within each slide.

Slides are then ordered and presented based on the size of the tumor regions identified.

This allows for an optimized, metric based review, where slides with most tumor can be assessed first.

The metastasis detection app is developed based on a deep neural network.

The principle is that the deep neural network learns a set of rules on how to recognize metastatic areas from normal lymph node tissue, tissue, based on a set of example images.

This approach is very similar to how pathologists work.

Pathologists are also presented with many examples of normal and metastatic tissue, and based on that, they develop their set of rules.

Because the network rules are taught from example images, it becomes extremely important to choose a good set of training images.

For this development, we made sure to include example images from five different clinical sites to assure a high amount of variance with respect to fixation, staining protocol, scanning, and so on.

All data was furthermore reviewed by skilled pathologists to make sure that the data quality was high. We have conducted clinical studies for this application.

And now I will give the word over to Janice, who has been heading the development and validation of the app, to present those results.

Thank you, Stine. So as Stine mentioned, I’ll be presenting the results from our validation studies and basically how we CE mark the metastasis detection app.

So in the validation studies, we wanted to investigate how a pathologist performs when using the app as decision support compared to conventional lymph node assessment.

So basically, how does the performance differ when scoring slides as negative or positive without and with the app. In the conventional workflow, the pathologists are presented with lymph nodes slides such as the one shown here. In the app assisted workflow, this type of lymph node images are analyzed before they’re presented to the pathologist and the pathologist will then instead be met with this image.

So in this image, the app has outlined potential metastatic areas.

These outlines quickly guides the pathologist’s attention to the most interesting areas on the image.

Just to give a small introduction to how we actually designed our validation studies.

In the validation studies, we had five different pathologists assess lymph node slides from breast cancer patients or colorectal cancer patients.

In the studies, the pathologists scored slides as positive or negative for metastasis.

Firstly they scored the slides without the app and afterwards with the app.

A washout period of at least four weeks between manual and app assisted scoring was used to eliminate memory bias.

What we tested was how a pathologist performs when using the app as decision support compared to manual lymph node assessment.

We used lymph node samples from breast and colorectal cancer patients as said.

This validation study design was used to generate the following initial results.

So the validation studies showed an agreement between manual and app assisted assessment in most of the approximately six hundred cases in the study.

However, we also found fourteen discrepant cases across the breast and colorectal cancer studies.

Together with the pathologist involved in the study, we investigated the discrepant cases further with a serial cytokeratin stained slide.

The cytokeratin stained slide was used by the pathologist to assess if the samples were truly negative or positive for metastasis.

Our discrepancy analysis revealed both the strengths and some limitations of the app and I will go through those findings now.

So seven of the fourteen cases turned out to be cases where the app helped catch a misdiagnosis.

I’ll just show you a few of such examples.

So here you see an H and E stained lymph node section from a breast cancer patient.

This sample was in the manual assessment assessed as being negative for metastasis.

However, if you look at the app outlines in this image, the app did detect metastatic regions and guided the pathologist attention to these areas.

This led to the sample being assessed as positive in the app assisted assessment.

The serial cytokeratin stain slide also shown here confirmed that the slide was truly positive for cancer.

Here you see a second case where the manual assessment resulted in a negative score of the sample, but the app assisted workflow managed to capture the tumor area and thus the pathologist scored the slide as positive. The H and E stain slide on the right shows the app outline of metastatic regions and the cytokeratin stain slide to the left shows that the slide is truly positive for metastasis.

In the studies, we also noted some limitations of the app and I’ll just show you a few examples of these.

The first example relates to poor data quality. On this slide, you can see that the tumor area is blurry and out of focus.

Due to the blurry appearance of the tumor area, only some areas are detected as suspicious by the app instead of the entire tumor area.

As we see here, slide quality is critical for automatic assessment of slides, be that with an app or manually.

If a slide is poorly fixated or stained, it can be impossible for the pathologist to give a correct reading.

And this is also why we here in Visiopharm are supporting EQAs.

If we look at the cytokeratin stained slide, we see that there actually are metastases present in this blurry area and that the app only managed to pick up relatively little of dose. So that is a clear limitation.

Another limitation we came across is the example shown here.

Here you see an H and E stained slide with three lymph nodes and a piece of tissue that belongs to the primary tumor.

As you can see, the app outlined an area in the non lymphatic tissue in the top right corner.

If we look at the cytokeratin stain slide, you can see that the areas the app outlined also match with what is positive on the cytokeratin.

So the app does not differentiate between lymphatic tissue and other tissue types, which can be seen as a limitation.

However, it is our belief that the pathologist would prefer to be presented with everything that the app considers suspicious so that they can make the slide assessment fully informed.

So in the end, we have the following table that sums up the results from our validation studies.

Our original question was, how does a pathologist perform when using the app as decision support compared to current clinical practice?

Our main conclusion to this question is that the app assisted workflow shows an increased detection of metastasis compared to the manual workflow for both the breast and the colorectal cancer studies.

We even found seven positive cases that were missed by manual assessment but detected Thank you, Dzenita.

I really think that these results show the true value of combining expert pathologist knowledge with artificial intelligence. And with that, I would like to thank you all for listening to our talk and over to you Bettina.

Thanks everybody for joining and thanks Dzenita and Stine for giving this excellent presentation.