Deep Learning, and … Dentists?

Ever chipped a tooth? Or, worse, have one knocked out? Assuming you have, the inevitable result is a premature trip to the dentist, followed by the oh-so-entertaining process of dental restoration.

For those of you who haven’t been through this, your dentist basically clean outs the area, takes an impression (there might be 3-D scanning involved here if high-tech-wizardry is involved, but it might also just be a resin cast), makes and fits a “crown”, and then painstakingly shapes it into place, adjusting all the time.

The last bit can be remarkably painstaking — there is a lot of custom handiwork involved here . It is quite the PITA, since not only does it have to fit in — plausibly!— with your teeth, but it also has to map into your bite, chewing, gripping etc. Get it wrong, and apart from just looking weird, it can end up making chewing difficult, cause TMJ disorders, and much worse.

You’d think that CAD/3D-Printing could be used to avoid most of this PITA, and you’d be wrong — the problem is that a lot of the work involved in getting the crown JustRight™ is, well, skill. Specifying all of this in code is just not feasible, with there being just way too many rules — and exceptions to these rules! — to specify.

When you think about it though, this is exactly the type of space where Deep Learning shines, no? We have a very large number of features (rules!), and a lot of examples showing both pre and post conditions (impressions! finished crowns!). You’d think that we should be able to throw a neural network at this problem, and get back a model that can be used to predict crowns for new cases, right?

Well, you’d be right — this is exactly what Hwang et al. have done in their recent paper “Learning Beyond Human Expertise with Generative Models for Dental Restorations”. They used the existing impressions and crowns as the “ground-truth” for input and output respectively for a type of neural network called a Generative Adversarial Network (GAN), and, as you might expect, it worked extremely well.

/via https://arxiv.org/pdf/1804.00064.pdf

What’s particularly fascinating here is that the model actually ended up outperforming the technicians, when considering how much penetration of the crown into the opposing jaw is allowed, and what the contact point distribution is (these factor significantly into biting and chewing functionality). To quote — “The generated crowns not only reach similar morphology quality as human experts’ designs but support better functionality enabled by learning through statistical features”.

Expect to see this becoming standardized as part of your dentist’s toolkit not too far in the future!