0:00 And now we’re going to go inside. Yes, sir. All right. Take it away. Thank you very much, Jan. I I do love the CDFAM for two reasons. Number one, 20 minutes is perfect for the challenged among us or the perpetually behind. So that’s really great. I can talk about just about anything for 20 minutes. I don’t know about you guys. And the second thing I really like is the no questions because then I can say whatever I want to and then none of you guys can challenge me.

0:35 But afterwards, if you want to if you want to chat, I would love to. For those of you that know who I am or what I talk about a lot, I’ll go through that first and then we’re going to spend a lot of time on the problem that we’re trying to solve. I think that’s important to understand contextually what the problem is, especially in the med device space.

0:53 Why we want to use things like machine learning which is this nice you know cool buzzword to to do what we do. We’ll talk a little bit about how people are sort of changing the paradigm in the orthopedic device space. We’ll talk a little bit about the old ways that we’re trying to pull people from and then talk a little bit about the preliminary solution that we’ve developed that utilizes some machine learning to do both geometry processing as well as anatomical tagging for things like patient specific applications.

1:22 So, a little about me. For those of you maybe that listen to my last talk, I now upgraded from four children to five children, which is a pretty big milestone. I spent my entire career Yeah, there’s some claps awkward. Spent my entire career in, orthopedic devices. About 10 of those and out of the manufacturing. One of cool distinctions I have is, I’ve been able to to guide the design direction for 3D printed implants for half of the world’s top 10 spine companies.

1:52 So, if you are into that and you want to Google some 3D printed spine devices, then there’s a 50% chance that you’ll see something that I’ve touched on at some point in the last decade or so. So let’s go into the problem a little bit. Now this is probably no surprise to you when looking at people. But you know, out of all 8 billion plus people in the world, every one of us is a little bit different.

2:16 Even you know by the time we mature even an identical twins even though the genetics are the same no two one of their bony structures are going to be identical so that we’re talking about surface morphology of course and then we’re also talking about internal morphology of the bone so when you take a sideby-side comparison of just about every CT scan bone that I’ve ever seen you’re going to have some sort of anatomical differences in fact again it’s not a it’s not a maybe it’s a certainty.

2:44 So because of that issue how many of you guys want to guess the percentage of humans with muscularkeeletal issues in your entire life? Me if I I won’t embarrass anybody but if I ask for a show of hands not maybe you young guys yet of those people that wake up and experience daily or weekly some sort of back or joint pain. Yeah. I mean you can if you want to the the majority of people should probably raise their hands.

3:12 And again, if you’ve not been afflicted with that yet, just wait till your 30s or 40s plus, right? It happens early. That’s because humans as a species are really tough on their bodies, on their muscularkeeletal system. The percentage of individuals needing some sort of orthopedic surgery in their lifetime, greater than 50%. So again, the an orthopedic surgery is the most common surgery performed in the United States and you can check the numbers.

3:42 Instrumented spine procedures, about 1.5 million procedures per year, and about 1.3 million joint procedures, total joint replacements between hip hips and knees, and I’ve not added things like total ankles as well, which probably add another couple hundred thousand to that. So, you might be thinking to yourself, man, for the most common major surgical intervention, we we certainly have great tools and we’ve refined the process so that this thing is like is great, right?

4:08 You would be mistaken. In fact, I like to say that most orthopedic surgery nowadays in the modern age is nothing more than glorified carpentry. For those of you that are curious and aren’t squeamish about it, Google and look up a recent, hip or knee total knee replacement. You would be surprised at the hammering and souling that’s taking place, during the procedure. Now, and again, nothing necessarily wrong with this.

4:38 The problem is the the industry is still focused a lot on fitting round pegs into square square pegs and round holes. The problem is the round hole is your body and they’re cutting a square peg into you to place an implant. The reason why this is a problem is because of the issues that arise from surgical interventions in the orthopedic space. Yeah, we have infections but infections actually account for a very small majority of cases.

5:05 The problem is something called aseptic loosening. So aseptic loosening is a a thing that we know has existed for a long time. That’s a problem and it occurs you see here in you know over half of all hip cases. The problem that’s going to require revision surgery is because of something’s loosening. And now there’s various reasons for this which you can see here and read through. I’m not going to read through all of them, but the primary one we wanted to focus on that’s just been seen in the industry over years and years and years is the idea about microotion.

5:37 Bone loves motion, but it likes it in a particular strain range. When it hits that strain range, bone grows really fast and it remodels and it integrates into an implant. When it doesn’t when it’s either at the disuse or the overuse phase it either removes bone that’s bad which causes loosening or it creates this woven yucky type of bone which again also contributes towards the aseptic loosening.

6:04 So what have we tried to do as engineers to solve the problem? Well these expandable devices are kind of this this stop gap to basically say we don’t know what is going to happen in the anatomy. Yeah, we’ve got some imaging techniques we can kind of tell what’s going to happen. So, we want to create devices that go into the body and then can be adjusted in situ.

6:21 Which has been honestly really cool. The ability to go in say we don’t quite know how this is going to fit. You know, jack up the spacer either, you know, inferior, superiorly, medially, things like that, anterior, posterior. The the problem is a lot of these devices are finicky. They’re expensive. And at the end of the day, you still don’t get a perfectly conformal fit. How could you?

6:46 Again, you’re you’re dealing with a device that’s manufactured as, you know, maybe a a set, but it’s not matched to that particular patient. This is actually what the vast 99% of cases come in with this multiple trays, hundreds of implants come into the operating room. You put in one or two or or maybe eight if you’re doing some bone screws. The fact is that you have such a massive array of products coming in.

7:12 So, you guys have heard the thing about one sizefits-all, right? Well, that’s not what we do in the orthopedic industry. It’s 300 plus sizes fits most. How’s that for a slogan, you know, you go pitch that to Nike, you know, or something like that. 300 plus sizes fits most. And this is the problem is the fact that we say we bring all these sizes in, but we still have revision rates because of aseptic loosening.

7:33 So, that’s a problem, right? So, let’s talk about, you know, this this dawn of a new era utilizing this idea about personalized medicine. We’ve personalized everything else, right? We started to personalize things like shoes. We personalize you know clothes. We got personalized jewelry. Now let’s talk about personalized implants. It makes sense, right? The you know something that’s going to live in your body as a replacement for one of your body parts should work over its lifetime, right?

8:00 It should be specific to you because your body is so unique. So there have been companies that have started to tackle this from an early stage. Conformis, which has been which actually I think was recently acquired by Restore 3D, has been doing sort of this similarly where they don’t necessarily take a scan and design an implant. They basically take a database of scans and match your particular scan to a cleared size that’s been tested that fits best.

8:28 This company Carl’s Med is doing this in spinal surgery for things like deformity correction. So deformity in spine has the highest revision rates because it’s very very hard to correct a deformed spine especially later on in somebody’s life when all the musculature has sort of grown into that to to to accommodate that specific curvature. And again down here at the bottom I’d say that you know what we’re trying to promise is stiffness matching although in my opinion we’ve not necessarily gotten there yet.

8:57 So these are these are products which show huge promise in this idea about personalized medicine. So we’ve got advanced imaging now. CT is really really good. MRI is pretty good as far as imaging and clarity. We’ve got auto segmentation algorithms to be able to take those things and turn them into usable data. We’ve got slice to 3D conversion and we have advanced design programs. The the problem is you know rush clear imaging is very common.

9:26 Algorithmic failure is very common a lot of the times because of imaging problems and then bad meshes bad geometry. I won’t get into that. I was one of the early adopters of the the Illuminati triangle when Duann years and years ago was saying down with the STL. But the fact is we still deal with bad meshes and bad geometry constantly. Now one thing I will give a nod to is advanced design programs have come a long way and they’re actually what I would consider to be the most robust at this point as far as additive.

9:52 And I’ll give a shout out for Enttop because I think what they’re doing with computational design programs, you’ll hear from them later in this conference is is pretty cool and we utilized their tool for some of this work. So the old ways of taking this idea about manual segmentation. How many of you guys work have ever like supported a 3D model and dealt with meshes and like magics or other similar programs?

10:14 So I like to call it painting triangles. You know what I’m talking about, right? You paint all the triangles that are the surface that you want to support. Well, this is similar. The the current standard way of doing this in the industry is to paint voxels. And what you’ll see here is it becomes really hard slice by slice to understand where a bone stops and where another one starts, especially in the joint layers.

10:34 So, you end up spending an immense work per case. Some of the things that you may see coming out of like the Mayo Clinic, the Mayo Clinic is doing incredibly cool work for patient specific applications for like lifealtering surgeries. The problem is they’re most likely spending tons of engineering effort hours per case. So that’s something we’re trying to avoid, right? You know, the the the more patient specific stuff we can do, the more lives that we can change, right?

10:57 So let’s talk a little about like and that’s the that’s the background for the work. And I wanted to spend a long time on the problem just to show you guys like that’s the deacto. 99% of cases are done the old way. 300 plus implants come into a set. 298 of them leave. Two of them get implanted. So what we ended up doing as we started looking at the problem is realize that there’s actually a ton of people already doing work in the space.

11:23 So we wanted to leverage open-source CTMRI data sets which exist. We wanted to leverage open-source deep learning models that already exist and you can see here again I happy to provide slides of some of the things but basically the self-configuring model for medical imaging has already existed. People have done a ton of work on it. As you can see here, the model shown as well as the type.

11:49 The database is actually called verse segmentation. It was a large scale spinal vertebral model, a database of several hundred scans that has essentially been pre-trained. So, you’re already starting with a really good point. And then our verification models against essentially this database was from an internal repository with some work we’re doing with a a customer on this particular application. So to further the problem the just because you have the 3D models gets you 40% of the way there when you go to do a patient specific surgery.

12:21 Now why is this important? Because just cuz I have the bones in 3D space I still have to design the implant. But certain things like anatomical landmark tagging such as the the width or the depth of the segment become incredibly important because when you go to put an implant in on the posterior side of that vertebral body is a spinal cord. And surgeons, not only do you have to fit it, but the surgeons need to clean the disc space.

12:44 So they also don’t want to have to clean back too far or they risk running into the spinal cord with a sharp tool. That’s bad, right? And certain other things become important in designing these sort of things like where is the bone the stiffest. We want to be able to land the implant on spots that we know are going to be supported over the life of the individual.

13:00 So if you put it on too soft of a spot, you get what’s called subsidance. The implant pushes into the bone. That’s bad. Generally requires a revision surgery and huge problems for the patient down the line. So to take the the landmark tagging, we wanted to to look at it two different ways. The first is looking at it with some sort of an AI model which a lot of people have started to to work on 3D anatomical landmark tagging.

13:24 The other thing we wanted to be able to do which was kind of like the stop gap while we were waiting for the programming of the the ML model was to do template matching. And this was actually done inside of NTOP. So what we basically did was we took a a template a ground truth with a tagged model and then we took our various inputs and what we did is we basically looked at the difference between the two models.

13:44 So we were able to generate each of them as fields. We’re able to look at the difference in the fields and then we’re able to essentially reverse interpolate the ground truth model data points back to each model that we’re doing with an incredibly high degree of success. So you can see here that the anatomical landmarks which again have been tested on on 10 vertebral well 10 spines and counting do a really remarkably good job of finding all the right anatomical landmark tagging because when you get to again the space where you’re inside of the body in 3D, you want to be able to know every single thing about the body that you can.

14:21 That way you can build your anatomical axes. That way when you go to place your implant, it’s in the right spot every single time. You don’t have to worry about tweaking or adjusting or flipping or anything like that. And just to give you guys an idea again, I’ve worked in this this space for a very long time to for traditionally for an engineer to design a patient specific implant.

14:41 Let’s say that they’ve got the model of the spine. That’s usually between back and forth with a surgeon with a design team, things like that. It usually takes about anywhere from 6 to 12 effort hours of design. That’s that’s fairly significant when you start talking about how long you or how many cases that you would be able to do without hugely scaling up your engineering effort hours.

15:06 Utilizing these tagged landmark models, we’re able to build a sort of repeatable pipeline which is able to accept CT scans or MRI scans and go to a final optimized model in about 3 minutes. So just just again to give you an idea I mean that’s that’s the why we do the parameterized models right is in order to be able to drastically reduce it but I think most people are really surprised because at the end of the day surgeons in particular are extraordinarily conservative.

15:34 They’re not necessarily going to just accept these sorts of things because they’re cool, right? There’s a lot of risk associated with placing an implant into a body. So the drastic reduction in workflow allows to allows us to be able to turn through more of these which allows us to be able to verify and validate this stuff more which gives the surgeons a lot of additional confidence in and sorts of things.

15:56 And then it also allows us to do also some really other cool stuff like find the joint spaces for spine surgery allows us to find the density and the height and the width of the pedacles which are incredibly important for posterior screws. So, one of the things I want to be able to go as we close out is I look back to our one-sizefits-all. I don’t think one sizefits-all will ever make sense in in surgical applications because no one product will ever be able to morph or adapt to fit everybody.

16:24 But one of the things that we are looking towards in the future is being able to develop a singular workflow that will fit every person regardless of anatomy, regardless of age, regardless of bone density. And that’s the vision we’re currently going on. So, thank you very much for listening. I appreciate your attention and I’ll hand it back to you, Duann.