The Imitation Game

Dental Lab Products, Dental Lab Products February 2021, Volume 43, Issue 1

How artificial intelligence technology has evolved, how it is poised to influence the dental lab industry, and what is needed for consistent adaptation.

With Google, Siri, and Alexa there to answer basically every question, it’s easy to assume that artificial intelligence (AI) and machine learning are new inventions. However, AI actually traces back to the mythological Greek Talos, a giant bronze automaton created to defend Crete from invaders.1 Talos was a fictional creation, but over time, automatons and robots became real devices. Now in 2021, this technology doesn’t just perform tasks—it learns.

It can be a bit unnerving when Amazon pushes ads for a product we might have just mentioned or thought about a few hours before. But the AI behind that ability is something that dental labs can leverage to their benefit.

What’s Possible

Machine learning is a function of AI that gives systems the ability to learn and improve from experience. Its algorithms use statistics to find patterns in large amounts of data.2 These algorithms are responsible for most of AI’s more well-known uses, including streaming platforms such as Netflix, YouTube, and Spotify; search engines like Google; and social media feeds such as Facebook and Twitter.2 According to Lisa Aguirre, a product manager of Dental Solutions at Roland DGA, DGSHAPE Americas, this technology can assist labs in distinctive processes and dimensions, including expanding production and other capabilities.

“By utilizing digital technology, such as Roland DG’s DGSHAPE DWX-52DCi with automatic disc changer, labs can take advantage of 24-hour unattended milling,” Aguirre says. “With DGSHAPE’s exclusive DWINDEX monitoring software, laboratory owners and technicians can monitor production on any mobile device or from the location of their choice. In addition, with DWINDEX software, lab owners can monitor the status of ongoing jobs throughout the workday, either remotely or from the comfort of their desk.”

Machine learning and AI can help throughout the business, starting with how a case enters the lab to the hours and expertise spent on it, to case planning, says Lou Azzara, CEO of Dental Services Group.

“The first 8 hours are critical to the success of the case and the relationship of practice and lab,” he says. “Getting that right is of utmost importance. Consider the expertise and talent of a lab tech who has seen thousands, potentially tens of thousands, of case scenarios. With their experience, they can navigate and see beyond the information that is presented. But if you incorporate AI, now you can start to do that at a scale that is beyond the ability of a single person or group of people and at a resolution that is unquestioned.”

Tolerance, the analysis of determining how close something is to its market or by how much it will deviate, is usually about 99.9% perfect with AI and machine learning, Azzara explains. This kind of deep understanding allows labs to establish a sort of institutional memory—building trends and predictability,3 he says.

Machine learning can also assist with communication among stakeholders. As there is often a lot of back-and-forth communication among dental practices and laboratories, AI can make it easy to share accurate information.

“AI allows easy access of recorded metrics for proper planning and can give us insights before and as cases are planned. The value of this information for the lab-clinician relationship is immense,” Azzara says. “AI analysis on clinical scans for prep clearance, marginal integrity, debonding probability, and ‘restoration success grade’ would allow the lab to communicate issues with real measurements attached, creating a more efficient system and reducing unnecessary costs that could be passed on to practices.”

Similarly, office managers and staff spend a great deal of time trying to understand past deficiencies and improving processes, Azzara notes. AI can be used in the front office setting to predict what’s around the corner and improve the patient experience. But AI’s potential isn’t limited to only these aspects. Human resources management, operational management, sales, communication, and lab and/or practice relationship can be positively impacted by AI, leading to a more effective experience for patients, Azzara says.

AI and machine learning are constantly processing information, and the technology can essentially learn all of our habits and eventually predict our decisions. This is something Jack Marrano, CDT, Absolute Dental Services’ director of Signature Prosthetics, refers to it as “intelligent design.” This can prove to be a great asset to the lab industry, he says.

“The vision for it should be huge,” Marrano says. “My cell phone knows everything about me. My Alexa at home knows everything about me because they’re constantly learning. Anytime we enter a data point, it learns. That’s what we’re doing when we’re designing. Each time I design a crown from the initial proposal, I generally will do the same things.”

Machine learning appears to be the next phase in the workflow evolution. As the industry embraces digital workflows, machine learning takes those efforts to a new level. Ophir Tanz, CEO of Pearl, a dental AI software company, considers the next level a second part of the digital revolution, which is like adding a cognitive layer to digital dentistry.

“The way that I think about it is that you had a real revolution at one point in digital dentistry, where things started to become digitized and nonanalog,” he says. “And that has been great in many ways for the category. Now you have a huge amount of digitized dental data that [are] historical and being generated every day. It’s difficult to make use of [those] data in a real way because there’s so much of it.”

Tanz is also a board member of the Dental AI Council (DAIC), an organization that studies and promotes AI in the dental field. “One way to think of AI and machine learning and dentistry is the adding of a cognitive layer to the digital dentistry realm,” he says. “So, it’s kind of like the second part of that revolution. Now you’re able to do things that, historically, would have required some creativity or intelligence, but you’re able to do them in a more automatic fashion.”

AI and machine learning can not only help dental labs automate but can also allow for greater customization within workflows, Tanz says. Using AI gives lab technicians the ability to create designs from the ground up without using a template.

“Areas where we see activity with machine learning in the laboratory environment involve things like smile design and crown design—which is a really interesting one because you’re no longer beholden to just some set of templates that you’re customizing,” he says. “You can really use AI to design a crown restoration from the ground up with the ideal contacts, occlusion, shape, and anatomy.”

Achieving Success

Sure, AI and machine learning have been around for a few years. But how do companies tailor it to the dental lab? And can machines learn dental anatomy as well as an experienced technician can?

These questions are an inspiration to Sergei Azernikov, PhD, vice president of AI and Machine Learning at Glidewell Dental. His team wasn’t achieving the level of automation it wanted.

“So, we started looking into new ways to do it, and a machine learning initiative started taking off. One of our first questions was, ‘Can we learn dental anatomy using machine learning?’ Can a neural network recognize teeth, for example? We asked ourselves, ‘Can we train a machine to recognize individual teeth?’” Azernikov says.

Once the team at Glidewell had AlexNet, its AI system set up—in a move reminiscent of a scene out of a science fiction film—it challenged an experienced, seasoned technician to a friendly competition against the machine. The results weren’t surprising.

“We trained the AlexNet network, which was a standard architecture, to identify teeth,” Azernikov says. “Even though that person was very experienced, he did make some mistakes, but the machine was accurate 100% of the time. This gave us confidence that the AI model can learn dental anatomy, and as such, it can be applied to many different applications. That’s how we convinced the company to invest in AI.”

That was 5 years ago, and machine learning has since been deployed in various aspects through the company. According to Azernikov, Glidewell first implemented this technology to track cases as they come in. The company began to take images of every case that came through their doors in an effort to recognize the boxes that dentists send in, types of impressions, and more, all of which used to be done manually.

When it comes to design, they’ve developed a system that generates crown proposals with little or no touches, Azernikov continued, and now the team is focused on doing the same for bridges.

“On the manufacturing side, when a design is ready, we have our robotic systems that go further,” he explains. “They take the blank and put it into [the] machine. Once it’s milled, we have an automatic [quality assurance] system that checks scans of the crown and compares it to the initial design to make sure that what we milled is actually what we designed. Then when it goes out, we also check whether it’s the right crown that specifically goes to that patient because for digital cases, we don’t have a model to test the crown on.”

When dental labs are looked at from a traditional standpoint, it’s easy to see how machine learning helps the overall workflow. According to Azernikov, Glidewell’s handles more than 10,000 cases per day. That kind of volume is similar to that of a mass-production manufacturing facility.

“But the big difference is that every case is unique,” he adds. “Even 1 person doesn’t have 2 identical teeth. So, it’s really a classic case of what’s called ‘mass customization’ or personalization at a huge scale.

“Our systems must be flexible enough to cope with all this variability, but at the same time, it needs to be able to consistently generate high-quality outcomes for our customers and their patients. Those are the things that are difficult because you can be very consistent if you do the same thing. If you make the same car, it’s easier, because all the parts are the same. But here, every part is different. So, how do we stay consistent across the board with a low price and high quality? That’s challenging. And that’s exactly why patient-specific products are the perfect case for machine learning. We just started, frankly, scratching the surface of where we can use it.”

The Cloud

Once the latest and greatest device is on the market, it’s hard not to want the newest model. We’ve all upgraded our cell phones, only for a fresher, sleeker model with better features to debut days or weeks later—it’s frustrating, and not to mention expensive. But one of the upsides of machine learning in its current state is that it’s not cost-prohibitive. The majority of the computer functions happen on the cloud. This is good news for dental labs looking to add AI and machine learning to their processes without the exorbitant expense of brand-new technology.

“The vast majority of folks out there aren’t going to engage directly in trying to build these types of capabilities,” Tanz says. “That is, at least in our experience, no hardware required whatsoever. So, no change to your existing infrastructure and no actual physical equipment costs whatsoever. The software tends to be incredibly lightweight. The reason for that is that all of this analysis and all these calculations are done in the cloud. So, what happens is—and this is the case for videography and oral scans—you capture whatever the asset may be, it gets sent out to the cloud,” he continued. “It gets processed there on very, very powerful machines that are a scaling infrastructure, and the results are sent back. Really, it’s the upload time, and then whatever text data is received back, which is obviously minuscule, and that tends to be the entirety of the integration. It’s as lightweight as anything you’ll find and typically fairly simple to engage with now.”

What’s Needed

Despite how capable these machines are, it still requires a bit of investment from the lab—mainly time and commitment. As powerful as these tools are, they can’t do all of the work for you, Azzara says.

“First and foremost, they need to adopt the mindset and dedication to make the investment, in a financial sense, but also in a time investment,” he says. “You must dedicate a plan to incorporate the tools and help your organization get through the learning curve, which can feel like a slow start to progress, but you’ll find the impact will start to gain traction. The development is largely based on having a culture that is innovative in thought, dedicated [to] the day-to-day, but also building a pathway to a transformational future. The system’s components are centered around the capturing, sending, and receiving of digital impressions, and with that, we need a streamlined workflow that keeps these files in their native format as much as possible.”

Achilles’ Heel

Just like other technology, AI and machine learning are not perfect systems. One of the challenges these machines face is interoperability, or the capacity for different information systems, devices, and applications to exchange data in a coordinated manner.44

“You have different setups in different infrastructures, in different environments,” Tanz explains. “Those companies don’t have the legacy plugins to their systems in order to pull out data, or they don’t want to share it, or whatever the case may be. Oftentimes, what you’re dealing with in terms of hindrances or bottlenecks, are just legacy systems, a fragmented environment, and needing to make things work in the reality of those environments. That’s something that people often overlook when you have an industry that’s been around for a while with lots of different software players and lots of old software and competing philosophical views about what should be open and what’s proprietary.”

One of the tech’s other biggest obstacles is the growing adoption of chairside systems, Azerniknov says. As those systems become easier to use and more mature, other doctors will feel confident enough in using the technology.

“Among other things, doctors will feel more confident to produce restorations while the patient is in the chair. In order to stay competitive, dental labs will have to invest heavily in technology to improve the level of service and reduce costs,” he says. “For example, we were recently able to provide 3-day turnaround to our customers, due to technological advances in manufacturing processes.”

Another issue could be relying on the computers too much. It’s one thing if a computer were allowed to design restorations all on its own, but human technicians will still want to be involved, according to Marrano, especially when proposals are involved.

“If you have 10 technicians and you line them up, each one will tell you that their crown is better than the technician sitting next to them,” he says. “There may be truth to it, but a lot of it is, ‘This is how we do it,’ and nobody is better than that than the technicians. I’ve spent a lot of time consulting in many labs, and when you walk in, every lab is the same, but every lab is different,” Marrano continues. “They all have their own way that they do restorations or their own standard of what they believe is ideal. So, I don’t think it could be something that is generated by a computer. If the computer system could generate the most perfect automated proposal, every technician out there would change it because the design was not theirs.”

Marrano believes that AI must learn from individual technicians to understand a particular lab’s process and standards. Otherwise, it’s self-defeating.

“If they just put something up there, everybody’s going to say, ‘Oh yeah, but I could do it better,’ and then they’re going to end up changing it, which kind of defeats the purpose,” Marrano says. “If I started with a standard molar and gave it to 10 different designers, you would get 10 different variations of that same molar design. And I can tell you, there was probably nothing wrong with the proposal in the first place, but they have to tinker with it.”

Aguirre believes that AI’s benefits outweigh its disadvantages. Dental digital technology and AI go hand-in-hand.

“Digital technology has progressed significantly over the last few years, providing more efficient workflows, unattended milling, and the ability to monitor lab jobs and milling machine functions offsite,” she says.

Misconceptions

Although machine learning has taken hold in some labs, it hasn’t yet seen broader adoption. The notion of what machine learning can do may not be clear to some, which could explain this reluctance.

“One of the key reasons that we helped found the DAIC is because this technology is so fundamentally misunderstood, and this isn’t only for dental labs,” Tanz says. “It’s really across every industry. And that’s of no fault of anyone in particular. [Machine learning] is a term that has been used and abused by Hollywood forever. Often, when you think about AI, the first thing that comes to mind is killer robots…we’ve been trained to interpret this term in a certain way. There is very little understanding about what this technology actually is, what it can do, what it can’t do, and what it will be able to do in the future.”

In the effort to prove 3 things, the DAIC engages in independent academic research to5:

  • Better understand how and where AI will most significantly impact the dental industry.
  • Validate the technology’s performance and capabilities.
  • Answer fundamental questions related to AI’s role in the oral health care ecosystem of tomorrow.

The first question deals with maintaining levels of consistency and proficiency among practitioners and laboratory technicians, Tanz explains.

“We just released a study where 136 practitioners looked at the same FMX [full mouth x-ray] set. What you find is diagnoses across the board and treatment plans between $300 and $36,000 and everywhere in between. So, there’s a real consistency problem with the kind of care that is getting issued for the same patient,” he says. “If the same patient goes to 10 different dentists, they are very likely to get 10 different diagnoses and treatment plans. That’s a problem. We can do better with AI.”

Tanz says the next question deals with the efficacy of AI and how to validate it academically.

“What are case-driven examples of how AI is going to be implemented across these various categories of the industry?” he asks. “Let’s really think critically and talk to leaders in those categories about the kind of impact that’s going to have.”

Another issue is that labs may expect an AI solution to take care of every aspect of the lab on its own in the short-term, which is simply not the case, Azzara says.

“In the short-term, they expect the AI software and tools to ‘show up and work,’ and the misconception is that it is not going to take additional work,” he says. “However, long-term underestimating of the power can seismically influence the way we operate.”

A solution to reining in expectations is understanding how AI is trained, Azzara adds.

“We are working within an industry that has, historically, been a hands-on problem-solving group,” he says. “Digital has been in dentistry for a while, but it has mostly been placed into the hands of newly hired talent with a great knowledge of technology, but less knowledge of dentistry. Therefore, the disconnect from the experienced, hands-on employees who know the ins and outs of dental is often removed from the more technologically inclined, but inexperienced digital group.”

Aguirre says that there may also be a misconception about what is possible with AI. Many make the mistake of assuming that adopting or using AI workflows are more complicated than they are when some are fairly easy to use.

“On the contrary, today’s digital dental technology provides efficient, easy-to-use solutions that expand both laboratory capabilities and overall production capacity,” she says. “The DGSHAPE DWX-52DCi’s 6-disc changer provides 24-hour unattended production and remote monitoring with its exclusive DWXINDEX monitoring software.”

In the bigger picture, AI and machine learning can be a difficult concept to picture, even with all the technology that consumes our daily lives. In many ways, it still feels so far away.

“I think most people don’t give it a thought,” Marrano says. “It’s one of those things you hear Elon Musk talking about, something that is way above most of our heads. It’s something that we don’t really, in the laboratory, interface with as of right now. I definitely see it being a part of laboratories in the future, but right now, I don’t think it crosses technicians’ minds.”

Another reason for the slow adoption of AI is the attitude toward new technology throughout the industry—it’s essential for some, a perk for another, and unnecessary for others.

“The dental industry is…pretty conservative,” Azernikov says. “And if you look at the first step, in order to do machine learning, you need the digital workflow, and that’s a key. Even though intraoral scanning has been available for decades now, still I would say only 20, 30 percent of doctors have it. Most doctors still work with analog impressions. To me, the doctors are still very slow in adopting those technologies.”

Adoption

What will it take for machine learning to take off? Don’t be surprised if the newest generation of dentists and lab technicians are the ones to embrace and drive the future of dental AI and machine learning, Azernikov says.

“It’s much more safe and faster and more accurate, and it’s a generational thing,” he says. “As the new generation comes in, they’re more open to new technologies. They were born into a digital world, so it’s easier for them to adopt this.”

An important piece of the puzzle has been cloud computing. According to Azernikov, before Glidewell even considered machine learning, they migrated their infrastructure to the Cloud via Amazon Web Services (AWS). Up until that point, their data were spread across several locations and hard drives.

“It was very difficult to collect it, to make sense out of it, [and] to analyze it was practically impossible,” he says. “Once we accumulated everything in one place, now we have more than 10 million cases, basically, stored on AWS, and it’s very easy for us to crunch it, to go through it, to collect the cases, to train our models, all using the data. That’s a very important step for anyone who wants to get into that space. You cannot have the legacy infrastructure and jump directly from that to the AI era.”

But, as Azernikov points out, humans still play a massive role. He suggests starting with simply having the right mentality when approaching AI; don’t see it as a threat or view it as taking away your job, but view it as a powerful tool to help you be more productive.

One of the obstacles here is finding or training the right team to implement these processes, Azernikov notes. It can be challenging to compete with the tech sector in retaining engineering talent.

“It is extremely difficult in today’s environment because the Amazons and Googles are going after these engineers as well,” he says. “And you need them if you want to develop or even just deploy AI technologies. You still need to understand enough to effectively use it.”

Marrano is hopeful for an even more accessible future for AI and machine learning. He says there should be individual logins for designers; that way, the machine starts learning the hundreds of thousands of movements that a technician makes.

“If I sit down at one of the designer’s computers right now, I can see how many movements they did to design the crown,” he says. And, more than likely, if I watched that one designer for a time period of, let’s say over an hour, they [will] continue to make the same adjustments that they did before. There’s no reason for it…my iPhone can do it. It’s very, very simple. They just need to implement it. That’s what’s going to revolutionize the industry right now.

“Single-unit design is gone,” he continues. “The future is AI design.”

References

  1. Skashkevich A. Stanford researcher examines earliest concepts of artificial intelligence, robots in ancient myths. Stanford University. February 28, 2019. Accessed January 14, 2021. https://news.stanford.edu/2019/02/28/ancient-myths-reveal-early-fantasies-artificial-life/#:~:text=The%20myth%20describes%20Talos%20as,boulders%20at%20approaching%20enemy%20ships
  2. Hao K. What is machine learning? MIT Technology Review. November 17, 2018. Accessed January 14, 2021. https://www.technologyreview.com/2018/11/17/103781/what-is-machine-learning-we-drew-you-another-flowchart/
  3. Mayo G. What is tolerance analysis? Shield AI. April 16, 2019. Accessed January 14, 2021. https://www.shield.ai/content/2019/4/16/what-tolerance-analysis
  4. Interoperability in healthcare. Healthcare Information and Management Systems Society, Inc. Accessed January 14, 2021. https://www.himss.org/resources/interoperability-healthcare
  5. Our purpose. Dental AI Council. Accessed January 14, 2021. https://www.dentalaicouncil.org/mission
download issueDownload Issue : Dental Lab Products February 2021