Carrefour de l'actualité

Teeth... as designed by AI

01 February 2022 - Source : BLOG

 

Integrating machine learning algorithms into dental prosthesis design process could decrease design error frequency. (Photo : Laboratoire de recherche MAGNU)


Artificial intelligence (AI) could soon find applications.... in our mouths!? A   Polytechnique Montréal team is working on this seemingly odd innovation in order to reduce the risk of errors when making dental crowns. It's a solution they hope will save hours for patients, dentists, and dental technicians alike.

If you've never had to have a dental crown fitted - you'd better find some wood to knock on - but also keep reading this post, so you'll understand all the challenges faced by those who design that kind of artificial tooth.

When a tooth breaks, is too worn, or has been damaged by a cavity in a significant way, the dentist often suggests replacing the damaged portion with a crown. They then install this prosthesis on the remaining healthy part of the tooth in order to protect it for years to come.

However before carrying out the latter procedure, the dentist and the prosthetist must first create what amounts to a new tooth cover of sorts. It's a job that requires meticulousness, because each prosthesis is unique. While the challenge is mostly overcome on the first try, sometimes the work must be corrected for one reason or another, which results in lost time for all involved.

Machine learning algorithms could correct the problem, and a team from Polytechnique Montréal is figuring out how.

close to perfection

Professor François Guibault (Photo : Mario Tremblay)

Full Professor François Guibault (Department of Computer Engineering and Software Engineering) and his team at Polytechnique Montréal are specialists in geometric modeling and shape optimization, using artificial intelligence to improve computer tools.

After focusing on turbine modeling, he and his team are now interested in your teeth! The team's objective is to create a computer tool that will support the work of prosthetists and help prevent errors when designing crowns.

Dentists and prosthetists already rely on digital tools to limit poor fit as much as possible. After machining the tooth to remove damaged portions and give it a shape capable of accommodating a crown (called "preparation"), the dentist takes a dental impression which is then digitized or captures the dental impression using a 3D camera.

The prosthetist then works with this in silico model. It positions a virtual tooth in the designated location, then modifies its shape so that it fits correctly. Once the work is done, this virtual prosthesis becomes real, and is usually made of fired porcelain.

The procedure may seem routine, yet according to Professor Guibault the reality is quite different. "We have to take several constraints into account," he says. First, the crown must fit almost perfectly on the preparation, "to within a few tens of microns," says the engineer. It must also avoid resting on the walls of neighboring teeth, while providing a surface that complements that of the opposing teeth for efficient chewing.

Since some mouths are smaller or differently shaped, "We must also take into account the trajectory to be taken during the crown installation process," points out Professor Guibault.

A look at... gan neural networks

A damaged tooth will be machined by a dentist to transform it into a preparation, on which the crown will be placedes. (Image : Stephen Wolfram, licence CC BY-SA 4.0)

To create crown designs that are nearly perfect, Professor Guibault's team relies on neural networks called "antagonistic generators" (GANs) - a made-in-Montréal concept building on work by colleague Professor Yoshua Bengio's team (Department of Computer Science and Operational Research at the Université de Montréal and Scientific Director of MILA).

The approach makes it possible to generate images, sounds or videos that pass for the real thing. In fact, "deepfakes" use this same strategy.

The GANs achieve their goal by pitting two algorithms with opposite objectives against each other. The first network - the generator -automatically creates "solutions" such as images, for example. The second network - a discriminator - acts as a filter, and compares the images offered by the generator with those from a database filled with real images. If it determines that the generator's image differs too much from the others, it discards it and tells the first algorithm (the generator), which uses this information to avoid repeating its mistake. This back and forth ends when the generator succeeds in deceiving the discriminator, when it no longer sees any difference between the real images and the one proposed by the generator.
"Using this strategy, we teach the network to generate a shape that could replace the part of the tooth that has been removed," explains Professor Guibault.

The image obtained is in fact a cloud of points. Professor Guibault's team creates it using three calculation operations carried out in parallel, each offering a different degree of resolution. “We then resurface in post-production from those point clouds,” he says.

Automating tooth segmentation

A damaged tooth will be machined by a dentist to transform it into a preparation, on which the crown will be placed. (Image : Laboratoire de recherche MAGNU)

Professor Guibault's team trains its AI tools using images of a few hundred crowns and mouths, data from its partners, the companies Intellident Dentaire, associated with the KerenOR dental laboratory, as well as iMD Research, which specializes in development of technologies for the dentistry community.

In addition to working on crown design, the team is also looking to automate tooth labeling, a step that comes upstream in the process.

Also assisting with the task (which Professor Guibault hopes to automate using AI) are students from CÉGEP Édouard Montpetit, recruited by the Collegial Center for Technology Transfer in Pharmaceutical Sciences, located at CÉGEP John-Abbott.

The labeling work consists of associating keywords with each image by indicating, for example, the number of each of the teeth to be identified as well as those to be repaired. By training with a dataset that's already been labeled, the algorithms should eventually automatically perform this task for any new image presented to them.

This work seems simpler than it really is, says Professor Guibault. “We sometimes work with dental arches where there are only a few teeth left, or which come from patients living with a congenital defect,” he notes.

Such a tool will be useful for the preparation of crowns, and other uses as well. “Any treatment planning requires that we first have to segment the teeth,” explains the Associate Professor in the Department of GIGL. “In the future, this tool could help dentists with other procedures as well."


Learn more

Professor François Guibault expertise
Department of Computer Engineering and Software Engineering website
Mesh and Numerical Geometry Research Laboratory (MAGNU) website

Comments

Commenter

* champs obligatoire
CAPTCHA
This question is for testing whether or not you are a human visitor and to prevent automated spam submissions.

Suggested Reading

August 5, 2020
NEWS

Examining aerosol movement in order to better protect health care personnel

February 25, 2021
NEWS

NSERC Alliance Grants: Polytechnique Montréal ranks first, winning 24% of all funds awarded to Québec universities