This text is part of the special Research section
A team of researchers from Polytechnique Montréal uses artificial intelligence to reduce the margins of error in the fabrication of dental crowns.
Hyperfaking at the service of dentistry, such is the motto of François Guibault, full professor in the Department of Computer Engineering and Software Engineering at Polytechnique Montréal. For 18 months, the researcher has been using the same artificial intelligence (AI) technology that is used to tamper with videos to make dental crowns.
“Making a tooth is as much a science as it is an art,” he says. Each mouth has its history and each tooth is unique to the individual, by its shape, its color, its surface, its wear. If I took a tooth out of a catalog and put it in, it would look artificial. »
The technician who makes a crown starts from measurements and digital images, then, using various software, generates a 3D shape that a machine will then reproduce in porcelain. “We are looking for a method to further automate certain processes”, specifies François Guibault.
A matter of etiquette
To generate a tooth or any digital shape, we start from a cloud of digital points which we then connect into small triangles. The shape of the tooth is the sum of all these little triangles. “When you scan a tooth, you can have 40,000 points. The whole thing is to know which are the good ones, ”explains the professor.
Before being able to produce the shape, it is therefore necessary to label each point, which consists in giving it a coordinate in x, y, z. In the case of teeth, which sometimes touch each other, the system must also determine the address of each coordinate — in other words, which tooth it belongs to.
“To achieve this, it takes a lot of mouths,” reveals François Guibault. He explains that the data comes from the Kerenor dental laboratory, in Westmount, which has thousands of scans and crowns. This collaboration also involves the firm iMD Research, of Montreal, specialized in dental technology.
“We now know how to label each tooth with more than 98% reliability. To achieve this, two students from Cégep Édouard-Montpetit, recruited by the College Center for Technology Transfer in Pharmaceutical Sciences, located at Cégep John-Abbott, worked for many months to analyze the point cloud and make the quality control and eliminate “noise”, i.e. irrelevant or redundant points.
But why generate data on a mouth when only one tooth is needed? First, because the teeth are often missing parts that need to be replaced. But also because a tooth never works alone. “We generate a tooth in its context. The ideal for us is to have the two neighboring teeth and the three opposite teeth. »
Make your neurons work
The next step is to generate the shape to reproduce it. For that, it is necessary to reconnect all the points to generate surfaces which one adds. François Guibault uses so-called neural circuits. These are “generative adversarial networks” (or GANs, for the English acronym), a deep learning technique developed in Montreal in 2014 and which has revolutionized AI.
Deep learning consists in putting in opposition two functions of the brain: that which generates and that which discriminates. The generator (also called encoder) produces the shapes; the discriminator (called a decoder) determines whether it looks like the model or not. The system learns by rejecting what is bad and keeping what is good.
So that the system can learn, it must therefore be provided with a lot of data. And this is where the data from the Kerenor dental laboratory is so important. “We have reached several hundred mouths. The more we show him scans and corresponding shapes, the better it becomes. »
François Guibault hopes to release the prototype by the summer. “For us, this will be the first opportunity to be used and criticized. That’s how genius is. We seek constant improvement,” he says.
The research engineer, who modeled hydraulic turbines before becoming interested in dental crowns, is already seeing other applications for his work. His lab uses the same technologies for additive manufacturing, commonly known as 3D printing, for titanium aircraft parts. But he also sees a use for the manufacture of prostheses. “Instead of pulling you a shin from the catalog, we could make your shin for you. »
However, he insists that his current work in dentistry is not intended to deprive technicians with 20 years of experience of their livelihood, on the contrary.
“Yes, if we’re successful, we’ll be able to generate a shape in seconds, while the technician can do it in an hour using various software. But it will always take his art and his experience for the adjustments and his appreciation of the result, he underlines. The objective is not to replace it, but to give it one more tool that will shorten its intervention and validation time. »