A New Way to Diagnose Osteoporosis

Ultrasonic tomography has been applied in many fields, from geophysics to engineering, and now to medicine. In this episode, we talk with Aaron Chung-Jukko and Peter Huthwaite (Imperial College London) about their work to develop an ultrasonic tomography algorithm that can be used to assess fracture risk in bones, and, as a result, be used as a noninvasive way to diagnose osteoporosis.

Associated paper:
- Aaron Chung-Jukko and Peter Huthwaite. "Virtual initialised ray tomography: Towards contact-free realistic ultrasonic bone imaging." J. Acoust. Soc. Am. 158, 276-290 (2025). https://doi.org/10.1121/10.0036902

Read more from The Journal of the Acoustical Society of America (JASA).
Learn more about Acoustical Society of America Publications.

Music Credit: Min 2019 by minwbu from Pixabay. 

Transcript

ASA Publications (00:25)

According to the National Institute of Health, about 10 million people over the age of 50 are affected by osteoporosis in the United States. It makes sense then that we need non-invasive methods to diagnose osteoporosis. Today I'm talking with Aaron Chung-Jukko and Peter Huthwaite about one possible method, which they present in their JASA Article, “Virtual initialised ray tomography: Towards contact-free realistic ultrasonic bone imaging.” Thank you for taking the time to speak with me today. How are you?

Peter Huthwaite (00:51)

I'm great, thank you.

Aaron Chung-Jukko (00:52)

Good to be here.

ASA Publications (00:54)

Good to have you. So first, tell us a bit about your research backgrounds.

Aaron Chung-Jukko (00:58)

Basically, this was my PhD thesis was around osteoporosis and ultrasound ,and I did a master's thesis on ultrasound and composites and I was given a few projects in the research group. I really wanted to work on something meaningful, if not as sexy as drones flying and robots and all that. So I partnered with Peter and this what we ended up working on.

ASA Publications (01:19)

I mean, it's very useful research, so.

Peter Huthwaite (01:25)

Yeah, and yeah, I mean, I think it's all very useful. Yes. So I started in the MechEng department here at Imperial as an undergraduate, actually, same as Aaron. And then I stayed on for a PhD in the non-destructive testing group, which generally applies ultrasound to engineering components. But actually my project there, my PhD project was on medical imaging, breast ultrasound tomography specifically. But then I stayed on the group, became an academic. Focused more on some of the traditional NDT applications for a few years, lots on oil pipelines, developing various techniques and also some ultrasound simulation tools. And in more recent years, I have been picking up more back on the medical side, trying to develop that a little bit further. And this is obviously where the bone ultrasound stuff comes in and the project Aaron's been working on for his PhD.

ASA Publications (02:16)

Oh, very cool. It's funny to see this overlap between pipes and bones, but... 

Peter Huthwaite (02:22)

Yeah, there's a lot of similar physics there that we can kind of exploit and try and cross pollinate. I think that's a phrase, isn't it? To try and get ideas from one area to the other.

ASA Publications (02:33)

Yeah, right, right. Okay, so to get started, what is ultrasound tomography and how is it used?

Aaron Chung-Jukko (02:38)

I guess the simplest way to talk about it is it’s creating an image of something that actually usually you can't see or you're trying to get a image of different properties spread across a space. So you might be trying to look at, in the case of medical imaging, then you might be trying to look for a certain let's say stiffness or other properties. So in the case of bone, then we're trying to look at the strength through some correlated properties.

Peter Huthwaite (03:08)

Yeah, I think it's worth sort of being specific. So it's an imaging modality, but within imaging it's often something where you have measurements from all directions and then try and produce some sort of quantitative map from that. So that's the sort of specifics of tomography, I would say, to add on to what Aaron said.

ASA Publications (03:25)

Okay, okay. So then what is bent ray tomography and how does it differ from regular ultrasound tomography?

Peter Huthwaite (03:33)

When we think about tomography, what we're trying to do is we're trying to interpret the complex physics of what happens when the ultrasound or whatever modality we're using interacts with whatever it's interacting with, whatever we want to image. So regular, well, regular tomography, we often think about CT radiographic tomography, where it's all very straight lines, it's standard x-rays, so we simplify that as straight lines and that works very well. When it comes to ultrasound, having nice straight rays is not very common. We get interactions with the objects, it causes scattering, it also causes refraction. Refraction is essentially the bending of waves. If you think when you look into a swimming pool, it's what makes the bottom appear at a weird level. So with bent ray tomography, what we're trying to do is correct for that, correct for this refraction effect and take account of it when we're producing our image.

ASA Publications (04:31)

Okay, okay. So why is bent ray tomography preferable for diagnosing hip fractures to the current gold standard for bone diagnostics, dual x-ray absorptionometry. Absorptionmentry. Absorption--

Aaron Chung-Jukko (04:44)

Yes. Difficult to pronounce. We usually would call it DEXA, D-E-X-A, DEXA. 

ASA Publications (04:53)

Oh, right!

Aaron Chung-Jukko (04:54)

So yeah, much easier. So basically it all comes down to the fundamentally what type of, how should I say this? Whether you're using X-ray or ultrasound or another thing, let's say MRI, to measure something. So with X-ray, what you're trying to measure is actually bone minerals. And it's very good at measuring bone minerals. So the way that DEXA works is that you have two different energies of X-rays, so different wavelengths, and based on the difference between images that you get out of the two different wavelengths or energies, then you're able to distinguish, okay, how much of the particular bone mineral that we're looking for you have. However, the problem with this is a bit like trying to… Let's say you do the same thing for a sandcastle; you're trying to basically image the strength of a sandcastle by measuring how much sand there is, which is a very important part. But as people, let's say most of us have tried or can imagine a sandcastle… There are other things to its strength. So in the case of sandcastles, you actually need water. If you just have dry sand in the Sahara Desert, you can't really build a sandcastle from it. So there is things that are missing that model where you're just trying to use bone mineral density to predict strength. So the reason that ultrasound has been suggested to be a sort of promising and inexpensive and portable alternative to DEXA is because ultrasound actually, the speed of sound is very strongly linked to stiffness and stiffness is very strongly linked to strength. So if, and this might be a big if, if we can use ultrasound to sort of tease out the speed of sound from the object, in this case bone, then we might have a better shot at figuring out the bone strength.

ASA Publications (06:42)

That's so cool. Okay, interesting. So what challenges arise when using bent ray tomography for imaging bone?

Aaron Chung-Jukko (06:50)

Yeah, so exactly as Peter mentioned with the analogy of the swimming pool, the problem is that exactly the thing that makes it possible to image the bone, which is the fact that it refracts when there's an interface where the speed of sound changes, similar to how refraction occurs in swimming pool between the surface of the water ,and so from the air to the water, so from your muscles, let's say, to the bone, at that interface then there are changes in the direction of the ray, if we can call it a ray. And so this is all good and well actually, normally, but the other part of this is that there are issues with a ray-based model. So we are quite used to nowadays, talking even just in common non-science domains, of sound waves. And so you can see in the word “bent ray tomography” that we're assuming ultrasound to be a ray. And this usually works quite well, but basically the sort of assumption of changing from a wave to a ray results in certain errors in the predictions. And that assumption works in certain scenarios, but when the distance between where you're transmitting the ultrasound wave in and where you're receiving it, as that distance increases, then the error increases in terms of its treatment as a ray.

ASA Publications (08:11)

Okay.

Peter Huthwaite (08:12)

Yes, I might just add a little bit more about bone specifically. So what the big challenge with bone that I think Aaron touched on a little bit when it comes to the refraction and you get the waves bending, but actually it causes a lot of the waves to be reflected because you've got a very big, we call impedance mismatch between the materials in the bone and the surrounding soft tissue. And that means that we don't really get very much ultrasound actually traveling in. A large fraction of that gets reflected and also not much gets back out when we try and measure or interpret what's going on. So it's kind of double whammy in that sense getting it in. And even worse once we actually get in there it's really quite attenuative so a lot of the signal really disappears. So it's a bit of a kind of combination of all these things that makes bone with ultrasound really quite challenging to deal with.

ASA Publications (09:05)

Okay, okay. So in your paper, you propose a solution to some of the issues with bent ray tomography. What was your goal for this work?

Aaron Chung-Jukko (09:13)

So really it’s to overcome the, as Peter mentioned, the high contrast problem as well as the sort of distance problem. In the paper we call it extended range imaging problem. So, sort of, we start from actually the problem and not so much coming to it with a solution already. 

So basically, in osteoporosis, one of the biggest problems is hip fractures. You can also have problems with vertebral fractures. But when you have a hip fracture, then mortality rates in sort of one year—so that just means your likelihood of dying within one year of having that particular fracture—is greatly increased. And this is not just coincidental, oh you're generally getting worse in health. You can actually separate out the direct causation rather than correlation of the effect. You can also imagine how not having mobility, not being able to exercise, not being able to have self-independence has a lot of impact in people's lives. 

But specifically the hip fracture. So there's an artery with the proximal femoral neck, which is where thigh bone connects into your hip bone. And that's the actual location that you want to image. And the thing with bone is that it varies in strength quite significantly throughout your body. And it's also got a special feature that basically it’s use it or lose it. So when you do exercise, then the bone over time will strengthen and when you are not using, you're not standing up, you're sitting all day, obviously many factors, then your bone will change. And that's why it's very important to be imaging at the exact location where the fracture happens, because you could have very different even results between your left thigh bone for example and your right thigh bone. And so just looking at the exact location of the hip imaging then to image at this particular location. So really you cannot take out the high contrast part of the problem. You cannot take out the extended range part of the problem, the extended range part of the problem arising from the fact that you do not have direct access to the thigh bone. You have your thigh around it.

ASA Publications (11:20)

Right.

Aaron Chung-Jukko (11:22)

So, hence that's why we try to… we look at bent-ray tomography as being very promising, but then particularly the two challenges of high contrast and the extended range are the ones we set out to solve.

ASA Publications (11:35)

So you guys have sort of a solution to some of these problems that just mentioned, which is virtualized initialized ray tomography. So what is virtualized initialized ray tomography and how does it address the challenges of bent ray-tomography?

Aaron Chung-Jukko (11:49)

Yeah, so one of the things that we realized—actually, it was kind of an accidental discovery—was that we could actually focus quite well onto the surface of the bone, but particularly focus in the way of getting a virtual transducer on the bone. So imagine you just have a ring of transducers, transducers being like a mini speaker that transmits the waves that you want. We found out that we can actually use those speakers in a particular way to reproduce the effect of a speaker as if it was directly touching the bone. So that is the key to the extended range part of the problem. So if you could image, literally on the bone, then that unlocks a whole range of different… The problem becomes much, much simpler. But also, through that focusing and through basically virtually having access to the surface of the bone, then that means that you also solve some of the noise issues, or the lack of ultrasound waves entering into the bone or coming out of the bone, it all becomes a much better signal-to-noise ratio. So that solves also some of the high contrast problem. But another part of this is actually, by actually changing from a sort of very big ring of transducers to exactly on the bone, you can start making assumptions about what's inside and what's outside. So you can basically say, “Oh, inside the small ring of the transducers, it's very likely that the speed of sound is much higher than outside. “And actually you can use very simple maths and very quickly try to get a initial solution for an average value of speed of sound within the bone. And from that point onwards, and that's the initialization step, from that point onwards it's much easier then to iterate with bent-ray tomography to a much better solution.

ASA Publications (13:46)

So how do you create this virtual transducer that's directly on the bone? Like, how do you make that happen?

Aaron Chung-Jukko (13:55)

Yeah, so the short answer for those who know ultrasound, then it's just delay and sum with a bit of sort of adjustment for strength. But actually the other way to look at it is that we backwards engineered what does a wave actually do if a virtual transducer were to exist at that point? What would you need to send from the external transducers? So if you rewind time, such that there was a wave at that exact point where you want it, then what do you need to play at the little speakers in order to recreate that exact waveform in the front? There were some minor sort of fudge factors and different minor changes there, but we basically find that it's quite good even if it's not exactly accurate. So you can do quite a lot of different things. However, one caveat I must say is that we, in this paper, we assume that we have a sort of good approximation, or we know what the external tissue is like. So we know the approximate speed of sound of the, let's say it's muscle or whatever other soft tissue there is out there. And we do make a simplification there.

ASA Publications (15:07)

Okay, yeah, that was going to be my next question. I was like, but people have different amounts of muscle around their bones, so that would probably affect things… But that makes sense.

Peter Huthwaite (15:15)

I think that is something that we could get over. There are, yeah, tomography techniques and I think, yeah, various different methods that you can use for extracting that sort of information and assumptions that you can make. And essentially what we're doing in this solution that we've been developing is exploiting the fact that we know that there is going to be this big solid bone in the middle. And we've been developing our algorithm to account for that and basically use that prior knowledge that we're going to have that big solid mass there in order to try and develop a better solution.

ASA Publications (15:49)

Right. OK. So did you end up showing that virtual initialized rate tomography achieves increased resolution and robustness compared to bent-ray tomography?

Aaron Chung-Jukko (16:00)

So, actually, we did this in multiple different ways, but one of main ways we did it was by doing a simulation of the sort of data set that you would get out of the mini speakers and microphones around the bone and around the soft tissue. And we, for the same exact data, we applied bent-ray tomography on it, see what the result is. So, we actually used a smiley face as well as a bone phantom that has some different features in it. So we did it with bent-ray tomography and then we did it with virtual initialized ray tomography and compared the results.

Peter Huthwaite (16:39)

Yeah, it's worth adding in a little bit about the simulation capabilities here. So simulation techniques have come on massively over the last, I don’t know,  20, 25, 30 years, to the point where we can actually do really quite accurate full 3D simulations, which enable us to capture a lot of the behavior you see within tissue, both soft tissue and the bone. And this has been really helpful for this work. We haven't actually been able to apply this to any experimental data yet. This would be something a bit down the line. But the performance with our simulation capabilities has been, has kind of convinced us of the merit in pursuing that further. It's worth mentioning, so we developed this package called POGO in the group, which has been primarily for non-destructive testing applications. It's a finite element ultrasound simulation package and we've been developing that also over to the medical side as well, trying to capture the sort of slightly different physics, things that occur that occur in the medical area.

Aaron Chung-Jukko (17:42)

Yeah, so to further answer the question more specifically around the increase in resolution and robustness. So if you, and it's very hard to describe sort of images, but so if we looked at the smiley example. In the original, then you have one eye that's a higher speed of sound, one eye that's lower, and then the mouth, I believe, is lower speed of sound as well, within a bone-like… Well, anyway, the smiley is made of bone.  We can wrap our head around that later on… But, so if you looked at the result for the bent-ray tomography, then you find that the smiley is just barely visible and it's like very small and also very faded. And the eyes are like sort of very near the top of the head and the mouth is sort of near the chin and it doesn't really do anything. I think without actually showing you any of the images of the other results, you wouldn't even be able to tell that it really is a smiley. But if you look at the, I'll start calling it VIRT, virtual initialized ray tomography, because that's faster. So if you look at the VIRT result, then you would that see everything is there. It has eyes, it has a face.

It's a bit blurry, but actually it's way better and the sort of general values of the bone are all a lot more accurate. I would say nearly perfect. So if you looked at the bone result, which actually is simulating, as I said, the area that we really, really want to image, which the femoral neck. 

And caveats, I can give you quite a few, but this was a 2d simulation, and the area is actually 3D, it changes in shape, et cetera, et cetera. But if you look at the results for bent-ray tomography versus VIRT, then you see that the bent-ray tomography result doesn't even get the geometry right. So the bone phantom actually has a variation in thickness. So this corresponds to the lower part of it actually anatomically is thicker, normally, and you wouldn't be able to see that at all on the bent-ray tomography result, whereas you could see some of that within the VIRT result. And we also added, I wouldn't say random but a different gradient in the opposite direction in the phantom just to tease out what you can get from within the cortical wall, so that's the cancellous bone, which there are many debates about what constitutes bone strength, but we're not going to wade into those debates. We're just going to say, can we image them? And the answer is that we can, a lot better than bent-ray tomography, we can see some of the gradient. It's not perfect, but we can see some of that. So just sort of headline results, overcome much of the resolution issues. Even you can say, geometry issues that bent-ray tomography experiences when trying to image from afar. But also, actually, a brief note on the robustness then. We actually had to really try to reduce the resolution when trying to use bent-ray tomography because it just crashes and it's not built for such high-resolution imaging.

ASA Publications (20:43)

With VIRT, you get a smiley that looks like a smiley with bent-ray tomography, you get no smiley or it doesn't work at all. 

Aaron Chung-Jukko (20:52)

Indeed.

ASA Publications (20:53)

Okay, so you had some different imaging configurations that you worked with in the simulation to kind of see how things, to do some comparisons. How did the various imaging configurations you compared end up performing?

Aaron Chung-Jukko (21:04)

We can start here probably with what's called Init BRT in the paper, but really it's, I think of this as a best case scenario for bent-ray tomography. What we do here is we add back the initialization step, but that being a very specific version of it, in that we give the algorithm a starting point where it assumes that there are two domains within the whole problem. There’s the bone and then there's the surrounding. So this is, one can argue, similar to the information that we give VIRT. And what you find is that now the geometry gets much better because we gave it the geometry, but actually the results are still far, far worse than the results in VIRT. And also it takes way longer to achieve results, one order of magnitude longer for worse results. The other imaging configuration we ended up trying is actually really, really much thanks to POGO that we can do this. So you know how we talked about using virtual transducers, creating transducers on the bone surface. So in POGO, we were able to directly just say, okay, let's put a virtual transducer or a tranducer in POGO directly on the bone surface. And so we did a side-by-side comparison without the virtualization step. What actually can you do? What's the limitations to this technique intrinsically, if you had access to the bone surface? How well would bent ray tomography do? And so we compared that result with VIRT. So looking at the smiley, then it does perform a bit better. There's no doubt about that. But actually, interestingly enough, for the, what we call the Bone200 example in the paper, then, it gets the cortical wall quite a bit better, but the cancellous interior of the bone, actually,  the information is all lost there. So there's some complex discussion there around why that I think we won't go into today. But essentially, the headline is that we basically are approaching the theoretical limits of what you can do with bent-ray tomography based approaches. We have nearly identical results as if you had direct contact with the bone so your little speakers and little microphones are all touching the bone. That’s the level of clarity that we're able to deliver with VIRT. And the final imaging configuration is, we can argue, not an imaging configuration so much as we just added noise into VIRT and see how it performs. So in the case of the smiley, the noise does make the result a bit blurrier, but the sort of overall speed of sound, which, again, reminder that it sort of correlates to bone strength, which is what we're trying to measure here. The overall numbers are all much better than BRT or even the best-case scenario for bent-ray tomography. And whereas for the bone case, so this is the bone 200, the semi-realistic phantom bone, then we find that while the cortical sort of thickness and the sort of material property prediction or the velocity prediction is less clear than without noise. The interior is actually the gradient of the cancellous bone is very, very well reconstructed. So there is a bit of a trade-off there on the bone case because of the guided waves. But headline is that it works really well 20 decibels noise, which is very, very strong noise.

ASA Publications (24:39)

Okay. Well, you already kind of touched on this, but what are some of key takeaways from this work?

Aaron Chung-Jukko (24:43)

If I go sort of take one step back, then I would say that actually it's that one can solve difficult problems with breaking it down. So in the case of the particular problem that we're solving, then that it really was down to breaking down it into sub problems. So you've got bone, you've got your external region, which has different properties. So it's just trying to go from just reducing the problem from the entire thing with one method and trying to solve it with one method, to saying, okay, this method is good for this part. And then another method is good for this part. But also then, more specifically, then the conclusion is that ray-based methods, or ray-based tomography, can be improved by focusing onto the surface of the object, and actually, this is very promising. Obviously, with further work with osteoporosis or bone imaging, but also actually, perhaps there are many applications where currently, bent-ray tomography is being used, but you could actually apply VIRT to get higher resolution with basically no additional cost other than just writing the code for it.

ASA Publications (25:54)

Very cool, yeah.

Peter Huthwaite (25:55)

Yeah, I think I'd back that up. So if you've got a problem where you've got a very small region or a relatively small region where there's some interesting things you want to focus in on that area, VIRT will work for that. So we can apply the same principles for that.

ASA Publications (26:10)

Awesome. So whether it's a bone or a pipe, it could be useful. So what are the next steps for VIRT in this research?

Peter Huthwaite (26:17)

So in terms of the next steps, I'm continuing to grow my ultrasound interests. So the principles that talked about here, I'm looking to try and apply to probably areas outside the bone as well. And one thing that I'm actually kind of specifically looking at in the medical area is to try and develop things where we don't need to have a kind of full circle of transducers all the way around. So looking more at a traditional ultrasound or medical ultrasound transducer, which would just be a linear array, and trying to essentially extract similar sort of information with that restricted setup. So that's something I plan to try and develop a little bit more.

Aaron Chung-Jukko (26:57)

And for me, yeah, it would be very exciting to continue to see where VIRT goes and where VIRT gets applied. I'm currently taking a tiny step back from academia, but I'm still very excited to see if VIRT can be applied and the method can be developed further, perhaps in or outside academia.

ASA Publications (27:16)

Awesome, very cool. Well, I am intrigued how both of your future works go with VIRT and without VIRT. It is really cool hear how ultrasound tomography could be, you know, the next step in helping with osteoporosis diagnosis. Thank you both so much for taking the time to speak with me today and have a great day.

Aaron Chung-Jukko (27:36)

Thank you.

Peter Huthwaite (27:36)

Thank you. Actually, Kat, can I just add one extra thing? So, also we’ve talked a lot about osteoporosis today. So, I wouldn’t mind just mentioning, I've got a friend from Imperial, hopefully in medicine, who has his own podcast called Bone Up, where he talks about osteoporosis. So, if any of your listeners have an interest in the osteoporosis side of things, you could have a listen to his podcast as well. As I say, it's called Bone Up.

ASA Publications (28:00)

Yeah, awesome, Bone Up, okay. Thank you.