Why Most Radiology AI Fails

So I tell people a body like this just doesn't come by accident, right? So as I did my research, Hopper is really addressing some big issues in healthcare, right? How to make imaging data usable for AI innovation.

What was the core insight that led you to launch the company? Can you provide us some insight there?

Yeah, Tom, I've been thinking about this for a long time.

I started working in machine learning and AI in 2004, before people were calling it AI.

And very early on, we realized there's some fundamental issues.

If you think of a radiology interpretation process, right? Everybody assumes it's the detection of disease is the problem.

It's not, right? I mean, we've known since 1975 that an average radiologist can find a half a centimeter lung nodule on chest x-rays within 200 milliseconds.

and a small lesion on the mammogram, also within 200-250 milliseconds.

These studies were published in the 1970s, still in the film days, and then reinforced again with Dr. Elizabeth Kripinski's work in perception in the 90s aspect of it.

So if you can find this so fast, then what's the actual problem? The actual problem really becomes is, taking all the findings you're seeing in an image on a study with thousands and thousands of images, and then synthesizing them into a communication device to talk to the clinician, which we call report, right? Report is not just, you know, regurgitating what you're seeing on the image.

It actually is what we look at image, and then we, based on what we are seeing, prioritize what the clinician would know.

If the same patient comes with the same disease, but the study was ordered by a primary care physician versus an oncology surgeon, my report will be very different.

Because in one scenario, I'm responding to a primary care.

In another scenario, I'm telling an oncological surgeon what he needs to be worried about, right? So that nuance is what creates cognitive loss, what creates time.

People just assume...

You know, you start seeing things, you start seeing it.

That is not how it works.

When I'm reading a CT scan, you know, I start with the neck, but I'm not talking about the neck first.

Patient came with right blue accordion pain, which is down in the belly, right? So I'm going to see all the stuff.

but not talk about those things first.

Talk about the things that are being asked for first.

But I have to keep all this information in my brain and synthesize them cohesively and make sure I don't miss anything when I'm talking about it.

So that's where the cognitive load comes in.

That's where the burnout comes in.

In the early days, when you had less images, maybe detection was helping the rare things be missed.