1
00:00:03,000 --> 00:01:57,000
Welcome to the Think Data podcast brought to you in partnership with Mydataworks. If you want to stay up to date with the latest breakthroughs and trends in the world of data and artificial intelligence, and if you're curious about some of the strategies that companies and founders use to launch data and AI products, then you're in the right place. Our aim is to bring together a diverse lineup of fantastic guests from the founders, through to accomplished leaders and product owners at some of the most fascinating data and AI companies worldwide. They will each offer you their own unique insight into what it takes to launch and scale a great data business. Thanks for tuning in and I hope you enjoy the episode. Welcome to the Think Data podcast and on today's show I'm joined by Saadi Salahid. Sadi is the director of research and principal scientist at Overjet. They are a really, really interesting business. They are a series C AI powered dental platform, which has been spun out of MIT. They use cutting edge machine learning techniques, primarily computer vision to analyze dental x-rays. And this helps dentist insurers and patients, basically it helps them make more accurate and transparent. care decisions and in this world obviously where the cost of kind of insurance cost of treatment especially obviously in the US where we're having this conversation it's a it's a business I've seen come out of pretty much nowhere over the last kind of few years and it's really good to have you on certainly from a kind of a more technical standpoint to our usual guests so yeah great to have you on Sadi and do you mind bringing everyone up to speed with who you are and I know you've been at Overjet for a couple of years now but what what brought you to the mission ultimately
2
00:01:58,000 --> 00:03:12,000
Thank you, Alex. Thank you for having me in your podcast. So I am Sadeq Saleri. I am a principal scientist and director of research here at Overjet. I have been here for now more than four years. So Overjet is a healthcare AR company focused on improving oral health by bringing more objectivity and consistency into dentistry. So we built learning-based models that analyze digital radiographs and help clinicians detect conditions like cavity, bone loss, and other pathologies earlier and more accurately. The goal is really to support dentists in making better decisions and also to improve transparency and trust with patients and insurers. So when I was before, I was at Hyperfine. And when the opportunity came up and they showed me what was the mission and how we can impact the oral health of the larger scale population, I was like, OK, yeah, that's a no brainer. It fits and the need is there. So I just jumped on this opportunity.
3
00:03:13,000 --> 00:03:45,000
Yeah, interesting. And I know I mentioned that kind of on the introduction that for me. personally I've seen overjet four or five years ago I saw you come onto the market and then out of nowhere I've obviously become the lead or one of the leaders in kind of dental AI I know you weren't there necessarily from day one but in your kind of more recent couple years being in the company what's what kind of journey have you seen in terms of where the organization was to kind of where they are now in terms of maybe some of the some of the problems they're solving that's a great question
4
00:03:45,000 --> 00:03:48,000
four years ago and now are completely different.
5
00:03:48,000 --> 00:03:48,000
Yeah.
6
00:03:48,000 --> 00:05:01,000
And both on the technological side and also on the access to data and the way that we can train and make our AI models ready. Four years ago, we had limited number of data. We didn't have so many clinics that are in our platform. So we needed to have a smaller model. We needed to have... more different models, smaller models that are experts in very specific tasks. But as we move forward, as we have more data, as we had more noisy labels from those initial models that we had, we were able to move toward more foundational models because now we have millions of data, we have millions of noisy labels from previous models. And now you can, you can train your model you can create that foundational model that is generalized on more variety of data as well as generalizable on more variety of the tasks that it can do so it's just
7
00:05:02,000 --> 00:05:40,000
two two two different worlds when you have access to almost unlimited data now and before we didn't have access to that part interesting what surprised you about the last couple years in terms of joining the business i know obviously you touched on it really valid obviously where we are now in terms of ai evolution obviously the sheer volume of data that's coming through in the access that overjet now has for those kind of that data. It's obviously far different to kind of where you guys were four years ago. But what surprised you in terms of the last couple of years, maybe from more of a research perspective and applying ML techniques to imaging specifically within dental?
8
00:05:41,000 --> 00:09:45,000
So one thing that is really surprisingly hard on the dental x-ray side compared to the, for example, just as an example, chest x-ray, is the variability. on different dimensions that we have, that this variability is not there in the chest x-ray. Just to give you an example. Chest x-ray, most of the images are coming on the perfect setup because the position of the device is always fixed. The position of the patient, it's easier to acquire the chest x-ray than the dental x-ray. On the dental x-ray, you have the hygienist or the dentist, when they capture, when they acquire data, they have different image types. And for each of these image types, it's very hard to capture the data. And the position of that image capturing, that device, can directly affect on what you observe from that video object that is the tooth. That's one variability that we don't see on the chair. And also on chest X-ray, you have only one observation of your 3D object. Your 3D object is your chest. And you see one observation and then you do your evaluation, your processing on that one observation. And then on X-ray, for every single 3D object, I'm not even separating the teeth right now. I'm just saying that, okay, one tooth, one molar. For that one molar tooth, you have at least two to three different 2D observations. And now these observations, you want to fuse these observations somehow to make a decision on that one object. That's okay. Is there a decay there or not? And sometimes you can see that decay on one of those observations and you cannot see on the other observation. Which one is correct? That's another dimension. That's another dimension that adds complexity to this question. One other complexity. It still amazes me how much complexity is there that you don't see as an outsider. When you start working on them, you realize that, oh, wow, there's a lot of complexity. One other complexity is that on the chess x-ray, you have one object. On day 23, you have... At least 32 objects. If you have the mixed dentition, you can have in one mouth. You can have pediatric, adult. Those are very limited. Each tooth and 3D object have different representation in your dental x-ray. Molar teeth are completely different than the anterior teeth. And then... When you look at those images, something that can appear as a decay on the molar, it is not decay in anterior because the anterior teeth, they are just thinner. Which means that the photons are passing through them easily. So now you see more shade on those anterior. So now you have the variability between teeth as well. So now if you want to... have the model that generalizes on all of these objects for that application. Now you need to take care of those variability as well. Make sure that your training data have these variability. Your evaluation habits, your model sees all these variability before you say that, okay, this is a good model. Let's put it in the production and show it to the customer.
9
00:09:45,000 --> 00:10:32,000
Yeah. And it's interesting because you're obviously building, your models are being used on both sides, right? You've got dentists and obviously those, the provider groups. And I know that in healthcare specifically, I know you fall under the broader remit of healthcare, there's still a lot of issues around kind of trust and... kind of organizations or people actually listening and believing you know what the data is telling them what the model is saying how do you get around that with regards to maybe the education to your customers or maybe the the AI system that you're building because ultimately if people aren't trusting what you're saying or your models are saying then people won't use your products how are you building the system to make sure that trust is you kind of everyone believes in the system
10
00:10:32,000 --> 00:13:35,000
That's a very good question. And I think we are, at Overjet, we are very well positioned ourselves to answer this exact question. And the way that we are positioning ourselves is that every single decision and every single output that we are showing to the customer and to anyone, either the insurer or the tenant, it is interpretable and it is evidence-based. So I can show and point to that evidence, to that image, and say that this is the reason that I am saying that. The way that we solve this problem is that we are quantifying the conditions and the pathologies inside of the image. And then based on that quantification, we make a decision. For example, there is a claim in the insurance. And. we say that yeah that's a good that's good the claim is good you should you should pay the provider for example the claim is that okay i needed to put the crown on tools number two because there was a very bad decay on that in order for us to look at that claim and see that yeah that's a that's a good claim we quantify all these evidence like Okay, there is a decay, yes. The size of the decay based on the size of the tooth is this much because we already segmented that decay and we know the resolution of the dental leaf state. So now we know that and we call it DMS, decay missing feeling ratio of that abnormality around the coronal area of the tooth. And we calculate that ratio and we are saying, okay, that ratio. It's 50%. And sure, above 50%, the best treatment, one of the best treatments is putting the crown for that tooth. That's it. Now you have evidence. You already quantified the abnormality and the pathology. And you make that decision based on that evidence. It is not like one model, black box, input image. would yes no okay that's that's not a good model in the the regulated space that that model can can have 99 accuracy or precision or recall or whatever but then what is the evidence why i should why i should trust your blackboard yeah overjet is not doing that overjet is looking at the evidence, quantifying it, and then based on that. That quantification makes a decision. And everyone can look at that and say, oh, yeah, sure. Because this is the number,
11
00:13:35,000 --> 00:14:23,000
this is the number, good. Interesting. Yeah, and I noticed before our call and obviously doing the research into you guys, you've recently made that move from pure software, because you were obviously a pure software player when you first launched, but you've now launched a hardware product, which is big because a lot of AI companies will specifically stay in their software lane and they'll just become... become better and stronger and faster within software what what prompted the move to hardware and how does that kind of create a better overall product for your companies and also for your customers two main quests two main two main reasons two main reasons on that side one is the unity of the platform because you want to have the platform you want to have your dentist or hygienist take
12
00:14:23,000 --> 00:15:17,000
the photo look at that photo if it is good send it out on the cloud side and get all the all the analysis i can come back and then then then do the treatment make a decision and the rest of the workflow so right now if you don't if you don't use the hardware you need you need to you need to use one set of hardware and software for data visualization then come back go to the auto our software platform to look at the segmentation to look at the overlay to look at to make a decision all those things But now when we have the IRIS, our data acquisition platform, now you run Unite platform that does everything. But also, one other thing is that by having the access to that software, we are decreasing the variability of image quality.
13
00:15:17,000 --> 00:15:18,000
Okay.
14
00:15:19,000 --> 00:15:30,000
From our perspective, because right now we are supporting almost all the major ones. But we can control the image quality and we know that it is good.
15
00:15:31,000 --> 00:15:33,000
Because you're relying on theirs right now. Exactly.
16
00:15:34,000 --> 00:18:17,000
And not only us, but also... We just recently got the FDA clearance on this one on Friday, literally like a week ago. But also, if you remember, I mentioned that, OK, the image capture quality is very different because it's the chest, for example. It's not like that. The operator, the hygienist that is taking the image can affect directly on how good or bad is this image and how much. how useful this image is for the decision making, the treatment decision making. And we just recently got DFT Clearance to have the model device, the edge model on device, not the cloud-based, it's a very fast one, the millisecond model, the very small model, that you take the image and in millisecond, it looks at all the, look at the image that you acquire. and say, is it a good image or not? Wow. Did you capture all the things that is necessary for this type of image or not? Did you position your acquisition device correctly or not? And if not, and you make it and you see that they are, again, we don't ask them to recapture the image, we just flag it and they make a decision if they need to. have we captured the image or not. But then you have this feedback loop in real time. So you don't need to capture the image, send it to the platform, because the person who is capturing the image and the person who is reading the image are not the same. Hygienist captured the image, dentist reading. So now the patient has already left the chair of the data capturing. They are now on the dentist chair. Dentists look at the image, oh, that's not a good image. You didn't position your device in a correct way, so I cannot make a decision if there's a decay here or not. Now the patient needs to come back to the other chair, take the correct image, and come back. Now we have it in real time. We just look at that. You see that we want to make this flow uninterrupted. And with the current... ability that they can provide the value is there and we can we can we can provide that value to just have the seamless experience for the patient and the provider and the dentist interesting the whole of
17
00:18:17,000 --> 00:18:41,000
how have you seen that i think it's fascinating that you've made that move and ultimately you'll understand why and understand kind of the benefit it brings how have you ensured that dental practices kind of get on board with that journey because ultimately again it's another It's a new product. It's a new way of working. It's a new kind of potential change to their current work. What have been the kind of key factors that have allowed you to kind of make that change?
18
00:18:42,000 --> 00:18:55,000
One thing that I realized is that the user, if they see the value, they just jump. It's as easy as that.
19
00:18:55,000 --> 00:19:00,000
And is it time? Are we talking about time saving here? Are we just talking about efficiencies? Is that the value you're adding?
20
00:19:00,000 --> 00:19:15,000
It's the efficiency that we are, about the value that we are adding. The other main value that we are adding right now for the dental offices is the communication between the dentist and the patient.
21
00:19:15,000 --> 00:19:16,000
Fine.
22
00:19:16,000 --> 00:19:34,000
Before overjet. The dentist showed the black and white x-ray image of the patient. I said, okay, there is a, you see that shape? Very small, weird thing that nobody knows about that. That's decay. And it's going to ruin your day if we don't treat it right now.
23
00:19:34,000 --> 00:19:36,000
I'm going to charge you a thousand bucks to fix it.
24
00:19:37,000 --> 00:19:42,000
But now, but now they look at, oh, wow, it is highlighted by AI.
25
00:19:42,000 --> 00:19:42,000
Yeah.
26
00:19:42,000 --> 00:20:04,000
And it, oh yeah, something is there. So now it's very easy for the dentist to communicate. because now the patient doesn't need to have the dent to go to dental school for four years to understand what really the problem on my tooth now we understand that okay that's as a second reviewer ai as a second reviewer
27
00:20:06,000 --> 00:20:39,000
agree with me that okay there is something there and we need to treat it and if you don't treat it you are going to have a heart yeah that's really interesting and i know you you mentioned right at the beginning the kind of journey or current state of where you are versus four years ago is obviously light years apart but over the next kind of kind of three five ten years if you looked at kind of AI within dentistry based on where Overjet are currently. Where do you think is the biggest opportunity and what does this look like in the next few years?
28
00:20:40,000 --> 00:22:01,000
Yeah, one thing that was not feasible two years ago, a year ago, and now it is feasible and it's really interesting and I'm really excited about this opportunity and we are actively working on is these reasoning models. Again, if you remember at the very beginning of the discussion, I was talking about the different representation of the same object right we have different uh different image types of the molar and we can have the intraoral scans we can have the cvcd we can have the even intraoral images just the rgb images of the same tool and then if you show these images the all different modalities to the dentist dentist is not making a decision by looking at one of these representations, one of these observations. They look at all these observations and they reason to make a decision. Okay, I see this thing here on this image, I see the other way on that image. Now I know that these two objects are representing the same object, same thing. And now I reason that, okay, this is, that's why now I can say that this is a decay with this side and all of this.
29
00:22:01,000 --> 00:22:03,000
Yeah, yeah.
30
00:22:03,000 --> 00:23:37,000
They reasoned. Now with all the advancement in LLMs and the reasoning, now we can do that on the model side as well. Now the model, the AI model, it reasons, not just quantifying and say that, okay, I saw that decay on this image. I see this side of the decay on this image. I didn't see the decay on this image, but the truth was not complete. completely observable on this image, so I'm not going to put so much weight on it. Now I'm reasoning that. So these reasons, the model, the LLM model, the large language model can do that and then make a decision. It was not possible two years ago. Now it is possible. Now it's easy. And now you can communicate that with identity even better. Now you have, rather than just having a couple of images with annotation, now you have the assistant that reason. And now the dentist can have the second opinion and consult with, okay, I see that thing. Do you see that? I saw you forgot. That image also has something interesting for the tooth number or whatever because all the reasons that we can think about that. So these are the things that... I think it's a very interesting and those opportunities are really interesting. And I'm really excited about these opportunities.
31
00:23:38,000 --> 00:23:59,000
Yeah, it's a fascinating space. Am I right in thinking it's solely focused in the US? And at the moment, you know, is there plans to expand beyond that? Because I'm guessing dental generally across globally is typically more of a private sector anyway, in terms of needing insurance companies. And do you see that expansion happening globally? you can have It's focused primarily in the US.
32
00:24:00,000 --> 00:24:49,000
And on the global side, we already started having that. We already have several clinics in Canada. We already have several clinics in UK. And we are expanding toward that space as well. But also you need to think about that for every single expansion, you need to make sure that you are following the regulatory on both sides. So that's... because it's a regulated space, you need to make sure that, okay, on the United States, you need to have the FTQ. On Canada, you need to have something like that. On the UK, you need to have something like that. All these things, just make it a bit slower for the expansion, but we are already doing that. So we already have the Canada and we are expanding beyond that as well.
33
00:24:49,000 --> 00:25:09,000
Because it's typically quite long in the US for FDA approval for anything. It typically is a really long, drawn out process as well. But I'm guessing with you guys, because you're already there, when you're releasing new models, new features, functionality, the integration with hardware, it's becoming more a bit easier, I guess, to get the approval for the next level.
34
00:25:10,000 --> 00:25:14,000
It does, but it's different.
35
00:25:15,000 --> 00:25:15,000
It's not easy.
36
00:25:16,000 --> 00:25:55,000
Yeah, it's not. It's not. It's not easy. It is different per country. And their agreement with the FDA on the United States. For Canada, you need to do something. For Australia, you need to do something. For the UK, you need to do something different. But most of the time, you don't need to. to rerun the whole evaluation from scratch because you already documented all those evaluations for the fdk layer so you already done that so you just need to make sure that you are you are compliant and compliant with those regulations makes
37
00:25:55,000 --> 00:26:05,000
sense and final question what's next for you and your team in terms of i know as much as you can say but uh what can you what exciting things you were working on which will be coming to the market soon
38
00:26:07,000 --> 00:27:30,000
One thing that is coming to the market right now is the CBCT. CBCT is the Combine CT. It's a 3D version of that. All those things that we just talked about. Oh, interesting. It's a 3D version, like the Combine CT. We already got the clearance on one of them. For the CBCT, we are filing for another one for the pathologies. And that's the exciting one. That's the one that I think it has even the value is there because it's really hard for the dentist to communicate with the patient on the 2D and it's almost impossible for them to communicate with the patient on the 3D. It's a very weird image that you really need to have a trained eye to look at those images. But now with the overjet CVCT, now you can have very easy communication with the patient. And also all the all different measurements that right now are being done manually. For example, when you do the root canal, when you do the endodontic treatment, some of these measurements, they just do it manually. And now we are just doing it automatically. These are the next exciting, very near future exciting one that we are releasing. And I'm really excited about that.
39
00:27:30,000 --> 00:27:56,000
Amazing. Well, look, it's been really interesting to have you on Saturday. I know there's a huge amount of information come through and I'll make sure that everyone checks over to you. I'll be checking you out on LinkedIn as well. But thanks for coming on this morning. It's been really, really fascinating and obviously really, really interesting to see a business that I saw a few years back to where you guys are now and all this kind of new, exciting stuff coming out in this age of AI is exciting. So thanks. Yeah, thanks for coming on again.
40
00:27:57,000 --> 00:28:03,000
thank you alex thanks alex for having me and the podcast my pleasure thank you take care
00:00:03,000 --> 00:01:57,000
Welcome to the Think Data podcast brought to you in partnership with Mydataworks. If you want to stay up to date with the latest breakthroughs and trends in the world of data and artificial intelligence, and if you're curious about some of the strategies that companies and founders use to launch data and AI products, then you're in the right place. Our aim is to bring together a diverse lineup of fantastic guests from the founders, through to accomplished leaders and product owners at some of the most fascinating data and AI companies worldwide. They will each offer you their own unique insight into what it takes to launch and scale a great data business. Thanks for tuning in and I hope you enjoy the episode. Welcome to the Think Data podcast and on today's show I'm joined by Saadi Salahid. Sadi is the director of research and principal scientist at Overjet. They are a really, really interesting business. They are a series C AI powered dental platform, which has been spun out of MIT. They use cutting edge machine learning techniques, primarily computer vision to analyze dental x-rays. And this helps dentist insurers and patients, basically it helps them make more accurate and transparent. care decisions and in this world obviously where the cost of kind of insurance cost of treatment especially obviously in the US where we're having this conversation it's a it's a business I've seen come out of pretty much nowhere over the last kind of few years and it's really good to have you on certainly from a kind of a more technical standpoint to our usual guests so yeah great to have you on Sadi and do you mind bringing everyone up to speed with who you are and I know you've been at Overjet for a couple of years now but what what brought you to the mission ultimately
2
00:01:58,000 --> 00:03:12,000
Thank you, Alex. Thank you for having me in your podcast. So I am Sadeq Saleri. I am a principal scientist and director of research here at Overjet. I have been here for now more than four years. So Overjet is a healthcare AR company focused on improving oral health by bringing more objectivity and consistency into dentistry. So we built learning-based models that analyze digital radiographs and help clinicians detect conditions like cavity, bone loss, and other pathologies earlier and more accurately. The goal is really to support dentists in making better decisions and also to improve transparency and trust with patients and insurers. So when I was before, I was at Hyperfine. And when the opportunity came up and they showed me what was the mission and how we can impact the oral health of the larger scale population, I was like, OK, yeah, that's a no brainer. It fits and the need is there. So I just jumped on this opportunity.
3
00:03:13,000 --> 00:03:45,000
Yeah, interesting. And I know I mentioned that kind of on the introduction that for me. personally I've seen overjet four or five years ago I saw you come onto the market and then out of nowhere I've obviously become the lead or one of the leaders in kind of dental AI I know you weren't there necessarily from day one but in your kind of more recent couple years being in the company what's what kind of journey have you seen in terms of where the organization was to kind of where they are now in terms of maybe some of the some of the problems they're solving that's a great question
4
00:03:45,000 --> 00:03:48,000
four years ago and now are completely different.
5
00:03:48,000 --> 00:03:48,000
Yeah.
6
00:03:48,000 --> 00:05:01,000
And both on the technological side and also on the access to data and the way that we can train and make our AI models ready. Four years ago, we had limited number of data. We didn't have so many clinics that are in our platform. So we needed to have a smaller model. We needed to have... more different models, smaller models that are experts in very specific tasks. But as we move forward, as we have more data, as we had more noisy labels from those initial models that we had, we were able to move toward more foundational models because now we have millions of data, we have millions of noisy labels from previous models. And now you can, you can train your model you can create that foundational model that is generalized on more variety of data as well as generalizable on more variety of the tasks that it can do so it's just
7
00:05:02,000 --> 00:05:40,000
two two two different worlds when you have access to almost unlimited data now and before we didn't have access to that part interesting what surprised you about the last couple years in terms of joining the business i know obviously you touched on it really valid obviously where we are now in terms of ai evolution obviously the sheer volume of data that's coming through in the access that overjet now has for those kind of that data. It's obviously far different to kind of where you guys were four years ago. But what surprised you in terms of the last couple of years, maybe from more of a research perspective and applying ML techniques to imaging specifically within dental?
8
00:05:41,000 --> 00:09:45,000
So one thing that is really surprisingly hard on the dental x-ray side compared to the, for example, just as an example, chest x-ray, is the variability. on different dimensions that we have, that this variability is not there in the chest x-ray. Just to give you an example. Chest x-ray, most of the images are coming on the perfect setup because the position of the device is always fixed. The position of the patient, it's easier to acquire the chest x-ray than the dental x-ray. On the dental x-ray, you have the hygienist or the dentist, when they capture, when they acquire data, they have different image types. And for each of these image types, it's very hard to capture the data. And the position of that image capturing, that device, can directly affect on what you observe from that video object that is the tooth. That's one variability that we don't see on the chair. And also on chest X-ray, you have only one observation of your 3D object. Your 3D object is your chest. And you see one observation and then you do your evaluation, your processing on that one observation. And then on X-ray, for every single 3D object, I'm not even separating the teeth right now. I'm just saying that, okay, one tooth, one molar. For that one molar tooth, you have at least two to three different 2D observations. And now these observations, you want to fuse these observations somehow to make a decision on that one object. That's okay. Is there a decay there or not? And sometimes you can see that decay on one of those observations and you cannot see on the other observation. Which one is correct? That's another dimension. That's another dimension that adds complexity to this question. One other complexity. It still amazes me how much complexity is there that you don't see as an outsider. When you start working on them, you realize that, oh, wow, there's a lot of complexity. One other complexity is that on the chess x-ray, you have one object. On day 23, you have... At least 32 objects. If you have the mixed dentition, you can have in one mouth. You can have pediatric, adult. Those are very limited. Each tooth and 3D object have different representation in your dental x-ray. Molar teeth are completely different than the anterior teeth. And then... When you look at those images, something that can appear as a decay on the molar, it is not decay in anterior because the anterior teeth, they are just thinner. Which means that the photons are passing through them easily. So now you see more shade on those anterior. So now you have the variability between teeth as well. So now if you want to... have the model that generalizes on all of these objects for that application. Now you need to take care of those variability as well. Make sure that your training data have these variability. Your evaluation habits, your model sees all these variability before you say that, okay, this is a good model. Let's put it in the production and show it to the customer.
9
00:09:45,000 --> 00:10:32,000
Yeah. And it's interesting because you're obviously building, your models are being used on both sides, right? You've got dentists and obviously those, the provider groups. And I know that in healthcare specifically, I know you fall under the broader remit of healthcare, there's still a lot of issues around kind of trust and... kind of organizations or people actually listening and believing you know what the data is telling them what the model is saying how do you get around that with regards to maybe the education to your customers or maybe the the AI system that you're building because ultimately if people aren't trusting what you're saying or your models are saying then people won't use your products how are you building the system to make sure that trust is you kind of everyone believes in the system
10
00:10:32,000 --> 00:13:35,000
That's a very good question. And I think we are, at Overjet, we are very well positioned ourselves to answer this exact question. And the way that we are positioning ourselves is that every single decision and every single output that we are showing to the customer and to anyone, either the insurer or the tenant, it is interpretable and it is evidence-based. So I can show and point to that evidence, to that image, and say that this is the reason that I am saying that. The way that we solve this problem is that we are quantifying the conditions and the pathologies inside of the image. And then based on that quantification, we make a decision. For example, there is a claim in the insurance. And. we say that yeah that's a good that's good the claim is good you should you should pay the provider for example the claim is that okay i needed to put the crown on tools number two because there was a very bad decay on that in order for us to look at that claim and see that yeah that's a that's a good claim we quantify all these evidence like Okay, there is a decay, yes. The size of the decay based on the size of the tooth is this much because we already segmented that decay and we know the resolution of the dental leaf state. So now we know that and we call it DMS, decay missing feeling ratio of that abnormality around the coronal area of the tooth. And we calculate that ratio and we are saying, okay, that ratio. It's 50%. And sure, above 50%, the best treatment, one of the best treatments is putting the crown for that tooth. That's it. Now you have evidence. You already quantified the abnormality and the pathology. And you make that decision based on that evidence. It is not like one model, black box, input image. would yes no okay that's that's not a good model in the the regulated space that that model can can have 99 accuracy or precision or recall or whatever but then what is the evidence why i should why i should trust your blackboard yeah overjet is not doing that overjet is looking at the evidence, quantifying it, and then based on that. That quantification makes a decision. And everyone can look at that and say, oh, yeah, sure. Because this is the number,
11
00:13:35,000 --> 00:14:23,000
this is the number, good. Interesting. Yeah, and I noticed before our call and obviously doing the research into you guys, you've recently made that move from pure software, because you were obviously a pure software player when you first launched, but you've now launched a hardware product, which is big because a lot of AI companies will specifically stay in their software lane and they'll just become... become better and stronger and faster within software what what prompted the move to hardware and how does that kind of create a better overall product for your companies and also for your customers two main quests two main two main reasons two main reasons on that side one is the unity of the platform because you want to have the platform you want to have your dentist or hygienist take
12
00:14:23,000 --> 00:15:17,000
the photo look at that photo if it is good send it out on the cloud side and get all the all the analysis i can come back and then then then do the treatment make a decision and the rest of the workflow so right now if you don't if you don't use the hardware you need you need to you need to use one set of hardware and software for data visualization then come back go to the auto our software platform to look at the segmentation to look at the overlay to look at to make a decision all those things But now when we have the IRIS, our data acquisition platform, now you run Unite platform that does everything. But also, one other thing is that by having the access to that software, we are decreasing the variability of image quality.
13
00:15:17,000 --> 00:15:18,000
Okay.
14
00:15:19,000 --> 00:15:30,000
From our perspective, because right now we are supporting almost all the major ones. But we can control the image quality and we know that it is good.
15
00:15:31,000 --> 00:15:33,000
Because you're relying on theirs right now. Exactly.
16
00:15:34,000 --> 00:18:17,000
And not only us, but also... We just recently got the FDA clearance on this one on Friday, literally like a week ago. But also, if you remember, I mentioned that, OK, the image capture quality is very different because it's the chest, for example. It's not like that. The operator, the hygienist that is taking the image can affect directly on how good or bad is this image and how much. how useful this image is for the decision making, the treatment decision making. And we just recently got DFT Clearance to have the model device, the edge model on device, not the cloud-based, it's a very fast one, the millisecond model, the very small model, that you take the image and in millisecond, it looks at all the, look at the image that you acquire. and say, is it a good image or not? Wow. Did you capture all the things that is necessary for this type of image or not? Did you position your acquisition device correctly or not? And if not, and you make it and you see that they are, again, we don't ask them to recapture the image, we just flag it and they make a decision if they need to. have we captured the image or not. But then you have this feedback loop in real time. So you don't need to capture the image, send it to the platform, because the person who is capturing the image and the person who is reading the image are not the same. Hygienist captured the image, dentist reading. So now the patient has already left the chair of the data capturing. They are now on the dentist chair. Dentists look at the image, oh, that's not a good image. You didn't position your device in a correct way, so I cannot make a decision if there's a decay here or not. Now the patient needs to come back to the other chair, take the correct image, and come back. Now we have it in real time. We just look at that. You see that we want to make this flow uninterrupted. And with the current... ability that they can provide the value is there and we can we can we can provide that value to just have the seamless experience for the patient and the provider and the dentist interesting the whole of
17
00:18:17,000 --> 00:18:41,000
how have you seen that i think it's fascinating that you've made that move and ultimately you'll understand why and understand kind of the benefit it brings how have you ensured that dental practices kind of get on board with that journey because ultimately again it's another It's a new product. It's a new way of working. It's a new kind of potential change to their current work. What have been the kind of key factors that have allowed you to kind of make that change?
18
00:18:42,000 --> 00:18:55,000
One thing that I realized is that the user, if they see the value, they just jump. It's as easy as that.
19
00:18:55,000 --> 00:19:00,000
And is it time? Are we talking about time saving here? Are we just talking about efficiencies? Is that the value you're adding?
20
00:19:00,000 --> 00:19:15,000
It's the efficiency that we are, about the value that we are adding. The other main value that we are adding right now for the dental offices is the communication between the dentist and the patient.
21
00:19:15,000 --> 00:19:16,000
Fine.
22
00:19:16,000 --> 00:19:34,000
Before overjet. The dentist showed the black and white x-ray image of the patient. I said, okay, there is a, you see that shape? Very small, weird thing that nobody knows about that. That's decay. And it's going to ruin your day if we don't treat it right now.
23
00:19:34,000 --> 00:19:36,000
I'm going to charge you a thousand bucks to fix it.
24
00:19:37,000 --> 00:19:42,000
But now, but now they look at, oh, wow, it is highlighted by AI.
25
00:19:42,000 --> 00:19:42,000
Yeah.
26
00:19:42,000 --> 00:20:04,000
And it, oh yeah, something is there. So now it's very easy for the dentist to communicate. because now the patient doesn't need to have the dent to go to dental school for four years to understand what really the problem on my tooth now we understand that okay that's as a second reviewer ai as a second reviewer
27
00:20:06,000 --> 00:20:39,000
agree with me that okay there is something there and we need to treat it and if you don't treat it you are going to have a heart yeah that's really interesting and i know you you mentioned right at the beginning the kind of journey or current state of where you are versus four years ago is obviously light years apart but over the next kind of kind of three five ten years if you looked at kind of AI within dentistry based on where Overjet are currently. Where do you think is the biggest opportunity and what does this look like in the next few years?
28
00:20:40,000 --> 00:22:01,000
Yeah, one thing that was not feasible two years ago, a year ago, and now it is feasible and it's really interesting and I'm really excited about this opportunity and we are actively working on is these reasoning models. Again, if you remember at the very beginning of the discussion, I was talking about the different representation of the same object right we have different uh different image types of the molar and we can have the intraoral scans we can have the cvcd we can have the even intraoral images just the rgb images of the same tool and then if you show these images the all different modalities to the dentist dentist is not making a decision by looking at one of these representations, one of these observations. They look at all these observations and they reason to make a decision. Okay, I see this thing here on this image, I see the other way on that image. Now I know that these two objects are representing the same object, same thing. And now I reason that, okay, this is, that's why now I can say that this is a decay with this side and all of this.
29
00:22:01,000 --> 00:22:03,000
Yeah, yeah.
30
00:22:03,000 --> 00:23:37,000
They reasoned. Now with all the advancement in LLMs and the reasoning, now we can do that on the model side as well. Now the model, the AI model, it reasons, not just quantifying and say that, okay, I saw that decay on this image. I see this side of the decay on this image. I didn't see the decay on this image, but the truth was not complete. completely observable on this image, so I'm not going to put so much weight on it. Now I'm reasoning that. So these reasons, the model, the LLM model, the large language model can do that and then make a decision. It was not possible two years ago. Now it is possible. Now it's easy. And now you can communicate that with identity even better. Now you have, rather than just having a couple of images with annotation, now you have the assistant that reason. And now the dentist can have the second opinion and consult with, okay, I see that thing. Do you see that? I saw you forgot. That image also has something interesting for the tooth number or whatever because all the reasons that we can think about that. So these are the things that... I think it's a very interesting and those opportunities are really interesting. And I'm really excited about these opportunities.
31
00:23:38,000 --> 00:23:59,000
Yeah, it's a fascinating space. Am I right in thinking it's solely focused in the US? And at the moment, you know, is there plans to expand beyond that? Because I'm guessing dental generally across globally is typically more of a private sector anyway, in terms of needing insurance companies. And do you see that expansion happening globally? you can have It's focused primarily in the US.
32
00:24:00,000 --> 00:24:49,000
And on the global side, we already started having that. We already have several clinics in Canada. We already have several clinics in UK. And we are expanding toward that space as well. But also you need to think about that for every single expansion, you need to make sure that you are following the regulatory on both sides. So that's... because it's a regulated space, you need to make sure that, okay, on the United States, you need to have the FTQ. On Canada, you need to have something like that. On the UK, you need to have something like that. All these things, just make it a bit slower for the expansion, but we are already doing that. So we already have the Canada and we are expanding beyond that as well.
33
00:24:49,000 --> 00:25:09,000
Because it's typically quite long in the US for FDA approval for anything. It typically is a really long, drawn out process as well. But I'm guessing with you guys, because you're already there, when you're releasing new models, new features, functionality, the integration with hardware, it's becoming more a bit easier, I guess, to get the approval for the next level.
34
00:25:10,000 --> 00:25:14,000
It does, but it's different.
35
00:25:15,000 --> 00:25:15,000
It's not easy.
36
00:25:16,000 --> 00:25:55,000
Yeah, it's not. It's not. It's not easy. It is different per country. And their agreement with the FDA on the United States. For Canada, you need to do something. For Australia, you need to do something. For the UK, you need to do something different. But most of the time, you don't need to. to rerun the whole evaluation from scratch because you already documented all those evaluations for the fdk layer so you already done that so you just need to make sure that you are you are compliant and compliant with those regulations makes
37
00:25:55,000 --> 00:26:05,000
sense and final question what's next for you and your team in terms of i know as much as you can say but uh what can you what exciting things you were working on which will be coming to the market soon
38
00:26:07,000 --> 00:27:30,000
One thing that is coming to the market right now is the CBCT. CBCT is the Combine CT. It's a 3D version of that. All those things that we just talked about. Oh, interesting. It's a 3D version, like the Combine CT. We already got the clearance on one of them. For the CBCT, we are filing for another one for the pathologies. And that's the exciting one. That's the one that I think it has even the value is there because it's really hard for the dentist to communicate with the patient on the 2D and it's almost impossible for them to communicate with the patient on the 3D. It's a very weird image that you really need to have a trained eye to look at those images. But now with the overjet CVCT, now you can have very easy communication with the patient. And also all the all different measurements that right now are being done manually. For example, when you do the root canal, when you do the endodontic treatment, some of these measurements, they just do it manually. And now we are just doing it automatically. These are the next exciting, very near future exciting one that we are releasing. And I'm really excited about that.
39
00:27:30,000 --> 00:27:56,000
Amazing. Well, look, it's been really interesting to have you on Saturday. I know there's a huge amount of information come through and I'll make sure that everyone checks over to you. I'll be checking you out on LinkedIn as well. But thanks for coming on this morning. It's been really, really fascinating and obviously really, really interesting to see a business that I saw a few years back to where you guys are now and all this kind of new, exciting stuff coming out in this age of AI is exciting. So thanks. Yeah, thanks for coming on again.
40
00:27:57,000 --> 00:28:03,000
thank you alex thanks alex for having me and the podcast my pleasure thank you take care
