What is the purpose of the virtual patient?
Imagine you're a surgeon, you are going to operate on a jaw, and you want to print a 3D model of this jaw to prevent a surgical plate.
So, you can attach it, do this jaw and let's say, replace it. That's a use case. That means you need to 3D print such a model. And before you can 3D print such a model; you first need to digitally create it.
In a nutshell, it helps to reduce this to a fully automated process: where the design is uploaded, then we let the AI program run, check, do diversification, and it's done. So that's the purpose of virtual patients.
How many patients did you use in the clinical study?
I think with AI studies, there is just a lot of data involved. I think we had, for instance, when validating the AI for recognizing each individual tooth, we had 200 patients.
Another one was for an airway, we had around a hundred patients. But it's often the gate for AI studies because you want to show and have a certain type of robustness. Now we not only use our own data but quite a thorough literature review on it. So, on average I think maybe a hundred per type of AI at least, I would say.
Which parameters did you use to compare the result of the virtual patients?
So basically, you need to imagine you're comparing 3D models against each other.
We use the accuracy of course, which is always above 99%. But I think some metrics provide a bit more insight on how well the AI is performing, like for example what we call intersection over union, which basically shows what is the overlapping part over the union of the two models.
That is as well, often beyond 80%, which is classified within AIS as very good. Then I want to know clinically if I can use it, so I measure how much space there is in between these two volumes. This kind of deviation, on average is way below half a millimetre, often we are only looking at, let's say 0.2 thanks to the accuracy of the raw image that has been taken.
What type of software skills do you require in this discipline?
We have an effective team fully focused on AI. They are engineers and they're creating algorithms making sure they run properly on servers. So maybe a coding language is Python, a framework like PyTorch.
Very strong algorithmic thinking is needed, often these are mathematical engineers. We also have web developers working in JavaScript and working in frameworks like Vue or React.
We have let's say the entire stack of engineers, but I think the most special profiles, are the guys thinking out the algorithms and on the front end, given that we have 3D web applications, just the people that really know well, how to develop in 3D, which is something you just don't learn at university. In fact, we had to train everybody here in house.
How do you protect your intellectual property rights?
Very good question. One of the best things in the software business is going fast, this is your best protection because software is just very hard to protect. I think in Europe, it's not even legally allowed. And I'm not talking by the way about patents, it is not legally allowed.
Furthermore, the algorithm we had three months ago, that's already completely changed, updated, improved on a lot of points. So, then why protect something that is changing so fast?
Trade secrets, legal things are in place as well as contractual agreements of course.
Do you have clinical evidence?
Yes, as mentioned before we have conducted academic research for a class IIa device, for the past two years where we have extracted insightful data.
We are also working closely with the University of Leiden, and we get a lot of interesting data regarding clinical evidence that they gather and review from all over the world.
The biggest conclusions are that with AI, your results are way faster and more accurate compared to doing it manually and easier in a way, as working in 3D is often very hard cognitively.
