Matthew Kurian, MD, and Arturo Loaiza-Bonilla, MD, examine how AI-powered pathology assays could help predict chemotherapy benefit in breast cancer using a single H&E slide. They discuss the potential for faster treatment decisions alongside challenges involving AI transparency, clinical validation, and adoption in routine practice.

Transcript:

Dr. Kurian:  So that leads us into one of the other abstracts. Another company called Ataraxis, which is focused more on breast cancer, has a study they are retrospectively validating and presenting at ESMO on the SWOG 8814 validation study. They are basically trying to do the same thing we just talked about, using retrospective data to help show that their assay is able to estimate individualized chemotherapy benefit too.

So what do you think in terms of these assays? As we are very familiar, they use single H&E slides. They are able to give an answer within two days sometimes, or less than seven days. Does that worry you at all, or is that just the power of the technology we’re dealing with?

Because our Oncotypes and MammaPrints sometimes take three weeks to come back, right? Now you’re saying you can make the same decisions within a day or two. I think some clinicians are a little bit skeptical of the information, the black box, and what is really making that decision.

It includes some clinical variables, but everyone’s worried about this black box, right? How is it making that decision exactly?

Dr. Loaiza-Bonilla: Yeah, and that’s also the key part because all these new technologies are hard for us to explain, so the ‘explainability’ component is so key because you only accept what you understand, right?

If you say, “Well, look at this slide set here, this single H&E slide, and this was a responder versus a non-responder,” and you look at them and you just cannot really discern the difference between them, it comes down to the tiny pixels around a nucleus.

That is the key challenge that we have in terms of adoption and deployment, right? So we can have the best tool in the world right now, but if we don’t have appropriate explainability or the technical capabilities to really deploy it in front of our patients, then we’re going to have a challenge.

I can personally attest that I have very good friends who are pathologists, and there is a really mixed view of these tools. Also, technology adoption at hospitals is very heterogeneous. Some folks have the capability to do whole slide imaging uploads and manage their own foundation models.

And some others, they’re like, “No, I have a microscope here. What do I need to do?”

So then we’re asking, as oncologists, not only for the oncologist and the patient, but we need the pathologists, the institution, the data, the guidelines, and the payer to all come together and fully align around a common vision of the results that we’re seeing in these conferences.

So I think we’re still in the maturity phase of this alignment, and I think that’s going to be one of the biggest challenges. It’s going to be more practical than anything else, more than the science itself. Because I am a believer in the science 100%. I see that, and I know that the data is the data.

You can see these are responders or non-responders. You can compare, and you can do masking. Basically, you blind the vision model, and then you kind of test it, and it actually tells you the outcome of what happened to that patient by looking at certain features that are key.

But that takes people like you and me, who are strong believers in these technologies, and we still have to look into it.

So imagine someone who’s just really new to the topic. It’s going to be challenging to really start digesting it.