Ophir Tanz, Founder & CEO of Pearl – Interview Series
Ophir Tanz, is the founder and CEO of Pearl, a company that was founded on the notion that artificial intelligence can be the dental practitioner’s always-on assistant and the patient’s most trustworthy friend. Its founders have a uniquely personal connection to the dental industry’s intricacies, as well as the knowledge and education to actualize the full and practicable potential that AI has to offer.
What initially attracted you to Artificial Intelligence?
I’ve been interested in AI since I was in college. I saw a lot of opportunity there and that drove my ambition to apply it to creating new capabilities and commercial applications. In particular, I was interested in computer vision –– the field of artificial intelligence where we teach computers to see, process and understand the world in the same way the human brain does –– so after graduating I launched a company, GumGum, that focused on applying visual machine intelligence to build value in the digital media category. Though I understood the power of AI quite early, as I grew that company, I was struck by how advanced and practicable the field was becoming ––and became increasingly interested in broader applications of the technology.
Your first company GumGum that specialized in using AI in contextual advertising ended up being hugely successful, what do you attribute this success to?
I think what allowed GumGum to succeed to the degree it has was the emphasis that we placed on AI application and innovation. It’s primarily a digital advertising company, but even though we operated within the broader confines of that category, the work we did with AI wasn’t actually confined by the category. That meant that we were a technology company as much as an adtech company which created significant differentiation. Our AI-first mindset led us to innovate in areas outside of the natural confines of digital advertising––in sponsorship valuation and, of course, dental. Because we were never focused on being “just an advertising company” and we're constantly looking for ways to do better, GumGum was able to grow with us as our vision expanded and the underlying technology and AI field evolved.
Could you share the genesis story behind your new AI startup Pearl?
After starting GumGum and focusing on computer vision, I knew there was more we could do with the technology and was always on the lookout for new applications. Healthcare and radiology were of particular interest to me, and also represented clear applications of the type of machine learning GumGum was applying. We launched a dental division called GumGum Dental, which was the genesis of Pearl. I decided to spin the dental division off completely because I believed the opportunity warranted a standalone company. I guess you could say it was meant to be in some ways – my father was a dentist, and I grew up helping out at his practice, so transitioning to focus on the dental industry was a little bit of a homecoming for me. But it’s not as though my childhood connection to dentistry was the main impetus for my desire to lead Pearl as a new venture. I strongly believe computer vision and AI will transform dentistry and global healthcare, and I wanted to be able to give the attention to the project that I feel it deserves.
Could you discuss the computer vision and machine learning systems that are used to scan radiographic and 3D dental imagery?
Computer vision is a form of AI that teaches computers to “see” in much the same way that humans do. We feed large amounts of expert-annotated dental images data into a series of algorithms that are modeled on the neural networks in the human brain. By studying the annotated images, the network learns how to recognize dental pathologies of the kind that have been marked in the annotated images. This process is called ‘supervised learning’. By teaching a computer this way, it can learn to recognize images in non-literal ways. For example, it learns how to identify a partially obscured object or one that is viewable only from certain angles by absorbing thousands of different examples and building what’s essentially a computer’s version of a mental image of that object.
We taught our AI and machine learning algorithms by building a large collection of radiographs and worked with dentists and radiologists to label the images, then used those labeled images to teach the system to interpret new images. Now we have an AI that can point to potential issues that can be identified in radiographs and help dentists read patient radiographs more accurately and consistently.
For our 3D imagery systems, we use a similar approach but with different classes of algorithms. With 3D, the training can be more complex, because 3D images contain so much more data, which sometimes makes annotation more laborious. Of course, because there’s so much more data, once the system has been trained to interpret a 3D image, it’s actually able to be more precise in its findings. It’s essentially the same as when a human looks at a cone beam versus a traditional bitewing radiograph: We can see every little facet of the tooth in a cone-beam computed tomography (CBCT), but we can often only just make out certain basic tooth structures in a bitewing. AI faces the same challenge.
What type of information or diagnosis is revealed by this system?
Our radiologic AI system can detect a large array of pathologic and non-pathologic conditions, restorative features, and natural anatomy. Caries, bone loss measurement, periapical radiolucency, calculus, crowding, calculus, impaction, WPL, furcation, obturation, margin discrepancy––the list is too long to enumerate everything and it keeps growing. Many of these capabilities are included in Second Opinion, our real-time pathology detection aid currently available in Canada, Australia, Europe and several other territories, and most are applied in Practice Intelligence, our non-patient facing clinical intelligence solution, which is available to practices in the US and globally
What type of imagery data was the system trained on?
Our radiologic pathology detection system was trained on bitewing, periapical and pano radiographs, which are most common in dental diagnostics––the kinds of x-rays that you get at the dentist every two years or so, and as the need arises. Radiographic images are relatively easy to obtain in the dental field compared to other forms of medicine and more dental radiographs are captured annually than any other form of radiography. The expensive and time-consuming part is getting experts review and annotate the x-rays. We’ve compiled the world’s largest collection of labeled dental x-rays. This availability of radiographic data is part of what makes the dental field so ripe for disruption by AI.
What type of efficiency improvements and accuracy rates have been seen from the Pearl system compared to manual human review of imagery?
We have conducted several large studies across thousands of radiographs and hundreds of dentists to test the accuracy of our system, both as a standalone detection system and when used to aid dentists. We’ve looked at accuracy for each detection type as well as broadly across all detections supported by the system. There is variance in accuracy between individual detection classes with accuracy ranging from around 84-96 percent. On the whole, the system is correct just over 92 percent of the time. That’s quite good and the system continues to improve.
Of course, these absolute accuracy figures aren’t actually as indicative as is the relative accuracy of the system compared to human dentists. If human accuracy were 60%, an AI system that was accurate only 70% of the time would provide a considerable advantage to dentists using it. In the studies we’ve conducted that included a human standalone component, dentists range from 70-85%. There is significant variance between individual dentists, however, so there are certainly some dentists out there who are equally or more accurate than our system and a good percentage who are far less accurate. To evaluate the benefit of the system, what we want to see is an increase in accuracy for a dentist when using the system compared to that same dentist when not using it. Our studies show a clear benefit there.
Now that Second Opinion is being used in practices, we need to do more research looking at real-world impact. We’re starting to do that with the help of academic partners in Germany. Does it speed up patient visits? Does it facilitate better doctor-patient communication? Does it improve patient trust? Does it elevate case acceptance? We’re currently working to answer these questions. Eventually, we’d like to investigate the system’s impact on patient health outcomes, but that’s a longer-term project.
I should note that, because Practice Intelligence is partly an analytics tool that can assess practice-wide patient health characteristics and the diagnostic and treatment planning performance of practitioners, we actually do have some sense of how AI can impact patient care. It is not academic-style research, but we recently performed a study looking at production data from ten Practice Intelligence-enabled offices over a one-month period. Over that month, the system surfaced an average of over $84,000 per practice in potential missed treatment opportunity in past radiographs for patients with scheduled appointments in that period. For that $84,000 in potential opportunity surfaced, the practices were able to complete an average of $12,500 in restorative treatment and an additional $23,800 in specialty treatment. That boost is coming from treatment opportunities that were previously missed. Because it was completed, we can assume these treatments were necessary and should have been provided after the patients’ previous visits. This was an informal case-study, but it seems to clearly show that AI brings significant benefits, both to patients and to the practices who are using it.
What in your opinion is holding back the wider adoption of AI in dental clinics?
The reception has been overwhelmingly positive from dentists using Second Opinion abroad and the thousand-plus offices who’ve deployed Practice Intelligence in the US, so there’s a segment of the industry that already has a desire for broad AI integration in dentistry. But wider adoption requires wider awareness. AI is new in the dental field. When we started working on dental radiology as GumGum Dental, we were, to my knowledge, the only commercial enterprise engaged in the effort. That was five years ago. The first marketable solutions emerged in late 2019 and they were insurance and laboratory applications, not clinical applications. We launched Practice Intelligence in 2020 and Second Opinion entered the global market in September 2021. So as far as most dentists are concerned, AI is a novelty. They need to be introduced to it and taught what it can and cannot do. There are some misconceptions about AI that need to be overcome. Certain dentists may incline to see AI as a threat, for example. Those misconceptions will be resolved as dentists become better informed about its utility. The benefits of AI are fundamentally attractive – higher standard of care, better oral healthcare, stronger financial outcomes for practices – so I expect adoption to accelerate rapidly once AI literacy in dentistry reaches a critical mass.
What’s your vision for the future of dental care in 10 years?
As the dental industry continues to embrace digital transformation, I see dentists incorporating AI into most of the time-consuming tasks they perform daily – like charting, scheduling, operations, inventory management – so that they focus on patients rather than on the routine tasks that take them away from the work for which their skills are uniquely suited. We’ll see a higher standard of patient care across the board, lower costs and, ultimately, a larger industry bringing better oral health to more people around the world.
I will also be surprised if, within 15 years, AI hasn’t begun to clear a pathway toward effective predictive diagnostics and personalized treatment planning. Is this individual patient at higher risk for cavities based on their genetic profile, lifestyle, past diagnoses? Can we recommend a preventative approach that will reduce their need for an invasive treatment in the future? If they have caries now, based on what we know about their individual characteristics, do we need to proceed with restorative treatment now or can we delay with the expectation that a specific change in lifestyle or consumption will likely abate the progression of decay? With the support of AI, we should be able to answer these questions––and, while we’re at it, perhaps narrow the unnatural gulf between oral and systemic health that exists today.
Is there anything else that you would like to share about Pearl?
Experts have been promising that AI will provide better clinical outcomes and cost savings in the healthcare industry for over a decade. Many of these promises have not been realized. Dentistry is actually a little late to the AI game, but AI is advancing in dentistry much more rapidly than in other healthcare categories. Why?
Considering medicine through a commercial lens, dentistry is much more entrepreneurial than other forms of medicine. Dentistry is performed in lots of small, traditionally privately owned practices. Most other forms of medicine are managed by hospitals, which are generally large bureaucratic corporate institutions. Dental practices and hospitals both have the same desire to increase efficiency, improve patient outcomes, etc, but structurally hospitals are too slow moving and conservative to effectively integrate and capitalize on emergent technologies that satisfy those desires. Dental practices, on the other hand, are agile––and the entrepreneurial character of dentists makes dentistry a far more fertile ground for innovations like AI. If a dentist sees a potential benefit in something, they can immediately implement it. A hospital will not be able to act with that kind of unilateral decisiveness. There will be feasibility and impact studies, pushback from countervailing interests and stakeholders, budget negotiations, and a gauntlet of other hoops through which a new technology will have to jump prior to implementation.
Equally important, however, is the fact that dentists can contribute to the effort to develop and improve it if they wish. Pearl has been able to conceive, build and commercialize this technology as rapidly as we have both because dentists are active and empowered consumers – we are developing products for a market that is unencumbered by the bureaucratic friction faced by companies trying to sell into hospitals – and because dentists are at liberty to put their material and intellectual support behind our efforts. Ultimately, our AI is as smart as it is because it has been trained and honed by an army of smart dentists who believe in the technology and were at liberty to contribute to its creation.
Thank you for the great interview, readers who wish to learn more should visit Pearl.
- NVIDIA: From Chipmaker to Trillion-Dollar AI Powerhouse
- Laura Petrich, PhD Student in Robotics & Machine Learning – Interview Series
- Liquid Neural Networks: Definition, Applications, & Challenges
- Patrick M. Pilarski, Ph.D. Canada CIFAR AI Chair (Amii) – Interview Series
- AI Leaders Warn of ‘Risk of Extinction’