How straightening your digital twin’s teeth might help yours

Back in March, there was a flurry of articles in the media on straightening teeth with AI. The stories focused on the work being done by the University of Copenhagen and 3Shape on predicting orthodontic treatments using AI. One of the key contributors to the university’s research was Ph.D. Torkan Gholamalizadeh. Torkan now works at 3Shape. She has a doctorate in computer science and works as a software developer in the machine learning and dental AI group at 3Shape. The team develops common AI components for 3Shape products in addition to focusing on university collaborations to advance the state of AI in dentistry.

Torkan took some time to talk with us about the University and her work.

Torkan Gholamalizadeh: “What the university and I worked on was segmenting teeth based on CBCT scans. In other words, we are segmenting the teeth down to the root level to create a “Digital Twin” of the patient. The twin could then be used to predict, for example, results of an orthodontic treatment.”   

The idea of segmenting teeth down to the root level is revolutionary. At present, 3Shape already segments teeth at the surface level with 3Shape Dental System and other software; however, segmenting anatomy down to the root level is a whole different ballgame.

Being able to do that would enable a much more accurate simulation of a treatment. Combining that segmentation with AI would allow AI to accurately predict tooth movement based on the forces applied to the teeth.

Torkan’s contribution to the work was in translating the CBCT images into segmented teeth and, eventually, into data because she is, of course, a computer science specialist. Her goal was to enable the computer to create a digital patient.

The challenges with segmenting teeth down to that level are enormous. After all, where does one tooth begin and the other start at the tissue level?  All our teeth have a connective tissue structure or periodontal ligaments (PDL) that surround each root and connect the root with the bone in the tooth socket.

While teeth are separated on the root level, the test becomes segmenting the tooth root from the bone structure. Bone structure is denser around your tooth sockets which can result in similar intensities on a CBCT image. If the quality of a CBCT scan is low, it becomes difficult to delineate the tooth-PDL-bone borders.

The question becomes, armed with that level of segmentation and computational modeling, could she accurately predict what braces or clear aligner treatment would do to the teeth?

It’s close, but Torkan says that there are more factors involved that also need to be accounted for.

For one, periodontal ligaments vary from person to person. Some are stiffer, with factors like age, alcohol, and smoking influencing the tissue’s elasticity. Another influence can be what is being used to move the teeth. If it is a clear aligner, Torkan found that every clear aligner maker uses different types of plastic with different thicknesses, which impact the teeth differently. In some cases, the clear aligner can become more elastic and, hence, less effective depending on the different design choices.

So, Torkan’s modeling needed to address not only the segmentation of the patient’s teeth down to the root level but also the health of that person’s tissue and even the manufacturer’s clear aligners that were being used for the treatment.

All these factors became parameters for her modeling. This is also why the concept of a digital twin has become so important. You have the anatomy of the patient captured in the digital twin, but the variables or parameters that need to be included are, for example, which clear aligner brand.

Meet your digital twin.

Back in 2023, 3Shape VP Rune Fisker spoke about digital twins in his 2023 trends in dentistry forecast blog post. [https://www.3shape.com/en/blog/2023/top-5-digital-dentistry-trends-2023]

In it, he said, “In dentistry, building a digital twin of a patient can mean combining the inputs from many different devices, sensors, and software.

This combined data will allow us to diagnose, plan, and treat patients at an entirely new level – mainly when adding in AI. I believe that we will use the digital twin for every treatment...”

His description and prediction are very much in line with Torkan’s work.

Torkan explains: “3Shape software works with intraoral scans. The segmentation is done on the crown level, meaning that we only segment what we can see in the patient’s mouth and nothing further.

When creating a digital twin, we include CBCT scans or x-ray scans that provide information on bone structure and the teeth roots. We segment the bone and teeth entirely (including crowns and roots), which helps us to create a digital twin that considers the entire anatomy of the teeth and bone.”

Torkan explained that there can be many challenges in mapping the tooth structure down to the root level. For example, the CBCT scans are not always sharp enough. A higher-resolution image would be better, but that would require a higher radiation dose to take the image, which is unacceptable for the patient. This was avoided by employing an AI model trained on accurately annotated CBCT scans in their research.

Torkan says that AI has helped because it can, for the most part, segment or create the actual geometry of the patient from the CBCT scans. AI can segment teeth based on a CBCT image in as little as two minutes – a task Torkan says would take an annotator a couple of days to do with a low quality CBCT scan.

AI in orthodontics

For now, AI modeling that she and the university have worked on has looked to predict the results of orthodontic treatment based on the patient’s anatomy (digital twin) and the braces or clear aligner manufacturer.

AI has accurately been able to predict that if a quantifiable amount of force or pressure is applied to a patient’s mouth, the teeth will move in a predictable way based on the patient’s anatomy - teeth and the periodontal ligaments - and which orthodontic device is applying the force.

The question for us became, can we turn this around and say, we want to achieve a specific outcome. Can AI tell us where and how to apply the force?

Torkan said that this would be the end goal of the research, but a framework for that has not been developed yet.

In this way, an orthodontist or dentist could choose what result they would like, and the AI would precisely identify where and what forces need to be applied to make it happen.

For the present, Torkan and the university’s solution should hopefully soon be a viable solution to help dental professionals predict the outcome of a treatment.

You can read Torkan's published clinical research here:

Deep-Learning-Based Segmentation of Individual Tooth and Bone with Periodontal Ligament Interface Details for Simulation Purposes - https://ieeexplore.ieee.org/document/10256039

Open-Full-Jaw: An open-access dataset and pipeline for finite element models of human jaw - https://www.sciencedirect.com/science/article/pii/S0169260722003911