Building an MRI scanner for under £5k - EMF 2026

[applause] Thanks guys. Uh yeah, so here's my talk which is building an MRI scanner for under 5K. And I thought uh the the best way to start this would be by taking uh a little 3D printed bit uh of plastic and putting a bit of butter in it of the EMF logo. And then you see the thing on the right is our MRI scan of that 3D printed uh part. So you can see there we're getting nice small details. Um you can see that sweepy line thing is about 0.5 mm which doesn't sound that good, but for an MRI scanner that's the sort of resolution you'd be getting if you go into a hospital to get one. Obviously, we're not built a hospital grade MRI scanner for under 5K. Uh, but we have got something which is getting decent images out. And this is the hardware we used to do it. I'll talk in a bit more detail about exactly what's involved in those different parts. We kind of got a journey of like stage one, we're getting an MRI signal to like better and better resolution and better pixels. So, by the end, we're scanning things like eggs and you can see quite a lot of detail on them. So, um, like how is this possible? Like I think when you talk to most people at MRIs, they're like, "How could you make an MRI for under 5K?" The ones in hospitals definitely don't cost under 5K. And I think this leads to a number of misconceptions about MRI, which makes sense why people have them, but they aren't fundamental to having a working MRI scanner. So, the first is you need a spinning magnet. Now, this kind of makes sense. If you go into a hospital, you see this like donut thing and you're like, "Oh, there must be something spinning inside it." There isn't anything spinning inside it. It's just a solenoid coil. There is a device which spins. It's called a CT scanner and that like spins a uh X-ray rounds and makes a bunch of different images. But MRI scanner doesn't need a spinning magnet. Perfectly stationary. Um the second one is you need a really strong magnet. So, if you know anything about MRI scanners, you'd be like, "Oh, if I have any metal on me, it'll like fly into the scanner and crush me." And that is true in a hospital one. Um, they normally have about 1.5 Tesla to three Tesla scanners, which is a big strong magnetic field, which could be dangerous, but you don't need a strong magnet. Strong magnets help. So, the stronger the magnetic field you have, the more signal you get, and so the better images you get. But it works for any magnetic field, regardless of how strong it is. In fact, people have done MRI on the earth's magnetic field. Uh but that's a story for another time. The final thing uh the final misconception is that you need superconductors. So uh there are certain things where you do need superconductors. So if you ever seen those cool things where you put a magnet on a floating track, that needs a superconductor to work. MRIs don't need superconducting. There's no kind of superconducting effect that's happening with them. It's just the only way that you can make a strong magnet is by having a coil of wire and putting a lot of current through it. And unless you've got a superconductor that gets really hot and would melt. So if you don't need these three things, what three things do you need? So you need three things. One is a uniform magnetic field. So the scanner we built was about 100 mila which is approximately 6% of the field strength of those um ones that you get in a hospital. And then you need something which makes small changes in that field. So the one we built, you've got 100 mila main field and then you change that field by about one millla from one part to the other. And if you ever been an MRI scanner, you'll know they're really noisy. That's that changing magnetic field. It's essentially MRI scanner kind of shaking itself. Um because when you change those magnetic fields, they're not infinitely rigid. So they will ride bait and that's what you hear as noise. And then the final thing you need is an RF transmitter and receive chain. And luckily for the thing that we're building about 100 milliter that hits you right into the amateur radio band van. So between five two to five mehz. So that's great because it means you've got lots of cheap radio amplifiers that you can use which is one of the reasons why you can make this so cheaply. It doesn't mean you have to do anything everything in a far cage because you've got all of the amateur radio stuff coming in uh and causing you noise and interference and you don't want to be causing noise and interference to all of the amateur radios around. So uh that's the three things you need. Um and I think before going into a bit more of what we've done, I'll give you like a very brief overview about how MRI works. So how to do MRI in five steps. Now before I say this, this is going to it's very hard to pitch it to all of you guys. Some of you guys will totally understand this and I will be explaining things you totally know. Some of you this will be too complicated and too detailed, but hopefully about 80% of you will be will be pitching it to the right level. come and find me afterwards if anything doesn't make sense or I've made a mistake or simplified something and we can talk a bit more about it. So the five steps. So you need to align, you need to resonate, you need to locate, you need to listen, and then you need to repeat. So let's start with align. We've got these nice little diagrams you can see. So if I go here. So the key fundamental physics reason why MRI works is if you put isotopes in a magnetic field, they will start um like imagine you've got a bunch of little tiny compasses. Uh they'll align with that magnetic field and start spinning and they'll start spinning at a certain frequency and the frequency is dependent on the isotope. So what we do with an MRI scan and what we've done with our scanner is we use hydrogen because in your body you have a lot of hydrogen because you have lots of water. So everywhere you've got fat, you've got bone, you've got muscle, it all has hydrogen isotopes in it. So you can pick up the resonant frequency of hydrogen for that. So what you do, you put a uniform magnetic field, they'll all start spinning or at least and not all of them. That's kind of like some of them will, some of them won't, but on average they'll all start spinning in the same direction. So the second thing you need to do is you need to resonate them. So this is kind of the magic trick of MRI is if you go too low a frequency and you put it in, nothing really happens. If you go too fast, you see the one at the bottom. I realize these diagrams are probably quite hard to see in this room in this light, but take it from me. Uh nothing really happens. But if you hear the right frequency, you'll get these like spinning compasses which will slowly get wider and wider. And what that means is like how magnetic fields are electromagnetic fields the two are linked hence this camp's name uh you can isolate that frequency and that means that it gets bigger and it will keep spinning which means you transmit something and then you can listen back at the same frequency. So then what you need to do is you need to locate. There are two ways to do this frequency and phase. So we'll start with frequency because I think it's the easiest to understand. So what you do is you've got let's have a really simple example. You've got three positions. Then what you do is you put what's called a readout gradient. So you'll have a weaker magnetic field on one side and a stronger magnetic field on the other side. Again the small numbers I was talking about. So your bulk field was about 100 milites. This is about like 1 mila difference from either side. But what that means is they will have different resonant frequencies. So the top we going slightly slower than the middle average and then the top one slightly faster. and you read out this sample at the bottom which is putting them all together. So you've got a bunch of signals of different frequencies coming down together. How do you separate them? Well, if you are uh an engineer of any form, you will have thought of fur transforms. So you can just do a simple f transform and that means everything from the first column you can get a f transform of the first frequency. Everything in the second column, second frequency, everything in the third column, third frequency. So that's frequency encoding. That's one of your axes. But images with like one-dimensional images aren't interesting. You want to have a two-dimensional image. So you have to do something for the different uh for the second axis. And what that is is phase encoding. Phase encoding is a bit more complicated. So bear with me. Hopefully I will explain it to at least some of you. So what we have with phase encoding is imagine you've got your top path and then your second path. You apply a gradient and you can offset it slightly. So what's going on here is imagine we're all on one frequency. All of that kind of frequency encoding thing is done which gives you your one vertical slice. Then what you can do is as I said your the the frequency at which it spins is proportion of magnetic field strength. So if you slightly increase it the one at the bottom you can make that one spin a little bit faster for a bit and then go back to the same frequency. And what that essentially does is it changes your phase. So if you can see you've got these two nice little diagrams here. So with one, they're both in sequence with each other. So you've got A plus B. The second one, we've moved B 180 degrees. So then you've got A minus B. And then it's pretty simple maths. You add the two together, the A's will add, the B's will cancel, you get A. Um, you subtract the two from each other. The A will cancel with the A and the B will cancel, the B will add with the minus B, which will give you B. So that's a really toy example for two lines, but you can extend it to get as many lines as you want. And it's kind of one of those things you had in school, kind of many equations, many unknowns. And then the fourth step, bear with me here, is you need to listen for it. So this wouldn't be particularly useful if you're getting the same signal back from different parts of the image. But luckily, different types of tissue have what a different what's called a T2 time. So they decay at different rates. So like you've got bone and you've got fat and you've got muscle. they will all kind of you know how was it was spinning around and it resonated it slowly goes back up so you don't get that signal forever but you get it back at different rates and those different rates mean that you can see different tissues so that's kind of the basic steps of MRI which is align resonate locate and listen but the final thing you need to do is repeat now I won't go through the math of this because I promise I won't go through the math of this but basically uh SNR is proportional to the square root of the number of repeats so it takes quite a a long time to like get really nice images if you're doing this in kind of a naive way because you have to keep doubling the amount of time that you're getting. And the reason why it works like this um is basically your signal adds up but your noise cancels itself. Um but it doesn't cancel itself perfectly. It cancels itself as a square root. So basically your signal is growing linearly, your noise is growing as a square root. So combine the two together and your SNR which is the amount of signal to noise ratio um grows with the square root of the number of repeats. So I'll go back to that slide I haven't been the one that we made for 5K and the various different parts that we had for it. So this is kind of like a fun journey of experimentation. Uh the one on the first is uh on the floor we had a bunch of different components. You've got like a RF radio amplifier. You've got a bunch of power supplies for things and the first magnet array we've got with a sample in it. And then with playing around a little bit, we managed to get our first MRI signal out of it, which was very exciting. Now, this was ignoring all of that gradient things I was talking about. It's literally just trying to find a resonant peak and get it back. And then we gradually went with more complicated designs. So, the first one we've got, we put it in this uh tool case kind of Faraday cage sort of thing. We got this first image which looking at now was pretty terrible but we were really excited about it because it showed that we got an image of something there. Um we then decided that magnet array wasn't working very well. So we built a new magnet array. I'll show you some pictures of that later. We got this nice V which is a nice image of things. And then we were like okay let's think about it. Let's try and refine our noise chain. Try and like make things better. And we made this nice rack mounted system. And then we've got like this picture of egg you can see here which isn't a chicken's egg. It's a quail's egg, which is much smaller. If you've never seen a quail's egg, just think of a very large mini egg, which I think is what they were based on. Um, the cool thing about it is you can see the egg yolk and the egg white pretty clearly. And also this little bit in the center, which I didn't really know eggs had beforehand, but apparently they've got a little bit of um going in the center to help feed the egg essentially. So, what made this so difficult? Well, the first thing is you kind of have to get everything right to get anything. So, when I was doing this project, there were lots of times where I'm like, it's amazing that anyone ever managed to make MRI work because you're kind of exploring the world of sorry, you're exploring the world of potential signals and uh if one part of your chain is broken, you just don't know it. You just don't get any signal. So you could be really close and then something was broken and you'll just debug something else and then that wasn't something you needed to debug. So once you've got like a working chain working, you can kind of hill climb from it. But until you get to that point, it's really hard to get there. And one of the reasons why it was quite hard to get there is actually really hard to know how strong a magnet is. So you can know approximately if you buy a magnetometer off the shelf or you do some simulations on your magnets. But knowing approximately how good it is isn't actually good enough for getting the spin escode sequences that you need for an MRI scanner. In fact, the magnetometers that you buy, the way they calibrate them is by putting them in like an NMR or an MRI scanner because that's the way that you can work out a magnetic field the most precisely. Um, you also have the problem that when you put your magnets together, they don't want to stay aligned. Um, so you might get a nice welltuned magnet and then it's kind of everything has slightly shifted, which means the magnetic field strength has changed because it has to be so precise. That's annoying. And the final thing is your magnets change magnetism with temperature, which is something that I didn't really know about beforehand, but like you run a scan and then everything will slightly heat up because you've got heat from your amplifier and things like that. And that will mean that you then get a slightly different magnetic field, which means when you're trying to get nice images and you're doing lots of repeats, you'll slowly drift off resonance and you have to compensate for that. So we had to do a lot of work to make sure that we were retuning uh as we went along to try and keep uh compensating for that change in temperature. And then the final thing is open source software is amazing. Like there were loads of open source packages we use. There's this big open source MRI project which we base lots of our stuff off, but it's also really confusing and difficult to debug and some of it's written in some languages that no one has used for ages. um or just kind of in a Python package that doesn't install. And this is one of the things where um I don't know if you're in the previous talk, AI is actually really great at this. So you can point AI at one of these repos and help um it'll help you debug it and like put it into a system that works for you. So this isn't like this isn't like saying anything bad about the open source software that we used. It was amazing and really grateful to people doing it, but also thanks like open a anthropic or the people who making AI to help that kind of distribute that kind of scientific knowledge into new projects. Um, so if you've ever had a kind of you found a GitHub repo that didn't work for something, it's actually really quite good to just throw AI at it and be like, can you rewrite this? Can you explain it? They generally can understand it quite well. So, uh, let's go through some photos of things we did. So, this is our first magnet array. You can see what we've got here is a h howback array of permanent magnets that we've printed into some 3D printed formers and then you've got a various coils and you've got a bit in the middle for hanging our sample. So the way that this works is you've got one magnet here and then you've got the other magnets that are rotating around. Basically the sideways parts cancel out and then you get a nice vertical field because of that. Uh, unfortunately, as I was saying, the magnets like twisting and 3D printed stuff is creeps slowly. So, you might get a nice magnet at first, but slowly it will get worse and worse as you go along with it. And the bit in the center, what we did was we printed these uh 3D printed things and put uh some jelly inside them. And that was where you got those first images out. You just kind of hang it in the center. Um, but this was kind of working, but we're like, "Oh, we need to redesign a new magnet array." So what we ended up with is we're like we'll print some magnetic discs. So 3D printers great thing if you want to make anything cheaply but including an MRI scanner. Maybe I made the talk we 3D printed an MRI scanner. Um so you here we've got our discs. So you have a disc of magnets. You print it out and then you push magnets into that and then you've got uh two plates here and you stick it on the plates and then this is our kind of magnet scanning rig to work out how uniform it is. So, uh, the nice thing about those Hback arrays, um, the circular ones, why people use them, is because they're nice and light because all of the flux is stored inside that loop. If you're using this, which is like more of a horseshoe magnet, you need some way of returning the flux from one side to the other. So, we've got these nice steel plates at the back of the front. And that's good and bad. It gives you a nice homogeneous field. It's a lot harder to simulate because if you've got a series of magnets, you can just add them up. Uh whereas if you've got something which isn't a magnet, it's a an iron yolk, you kind of have to do FEA simulation to work out how things are going through that. So we did that. We just kept adding steel until we weren't um saturating our yolk. Uh and then we mapped it. So what we did is we bought a cheap CNC machine off eBay. Uh and then we strapped a magnet to it. And then what we could do is we could get these nice nice little plots of our magnet homogeneity which I realize you won't be able to read any of the numbers on them. But essentially uh even color means that it's homogeneous. And by the one at the right we're about 250 ppm. So 250 parts per million of homogeneity which is what you need effectively to scan something. So it really is you need a very homogeneous field to get a nice nice image out. Um, the other thing we discovered was that if you get a power supply off the shelf, it's quite noisy. So, they're quite good for most applications. They're not good for when you need the nano volts coming out of an MRI scanner. So, we had to build our own battery power supply that could run the gradient cores and run the RF system um to uh make that work. And this was a fun like safety challenge to make sure that we were not going to set fire to anything, which we haven't, which is good. But this is the other thing. If you're building one of these systems, it's actually easier to run it off batteries than it is off the mains, just cuz you really want to reduce noise as many different sources as you can as possible. Uh these are the coils that we made. So this uh coil on the left, this is what an MRI scanner coil looks like. uh it's essentially just a solenoid, but you want it to be as thick as possible so that you reduce your resistance as much as possible so you get as much signal as possible. And then you've also got this shield that we used uh for some experiments which is to try and like you're in a Faraday cage but you've also got your gradients which will give you a bit of ant noise in antenna. So if you put a shield around your coil you'll also reduce the amount of noise that's there. Um, and these coils, one of the reasons we can make them work is we need a tuning circuit. So, we there's this thing in um RF called a Q factor, which is essentially a quality factor, and it's how much you amplify your signal. And what we really want to do is we run a really sharp Q factor, which amplifies the signal right at that frequency we care about and knocks out all of the noise apart from uh in different places. And so, we had to make up these fun little boards uh where you have to use your tuning and matching capacitors. One of the problems is for the tolerances you're talking about here, capacitors off the shelf don't work that well. So, um, well, they don't they work fine. It's just they're not the number they said. Like the tolerances are so large. So, you'll have to get a bunch of capacitors, try them, solder them on, get another one, solder them on, and then we've got these nice little turning things that you can rotate to get it really tuned on. It's one of the reasons why it was painful beforehand was because if you were off if you were off tuning, you just won't get signal. you have to make sure you're tuned at the right frequency, but until you know your actual magnetic field strength, you don't know what frequency to tune to. So, there's lots of like things where you have to get one thing right and get another thing right. Um, and yeah, we pushed this all together finally into a nice rack mounted system. So, you can see at the bottom we've got our battery pack. Then you've got uh what are our gradients and an uh AM uh amateur radio transmission set that uh basically the system that you've got is you've got a nice FTBA board which is generating your signals. Then you put that through a pre-amplifier. So we got some cheap mini circuit ones offline. Then we bought like a 300 uh radio amplifier which is doing your power amplification that goes into your transmit coils. You then have a switch uh which is a fun circuit which essentially means you can use one coil to transmit and receive and it switches between the two of them quickly. And the reason why you need to do that you can't just have one transmit and one receive is because your transmit power is so much greater than your receive power. You'd blow up your receive chain when you're transmitting. So you we use this uh open source project which was a transmit switch which essentially said now I'm transmitting now I'm receiving. So, which means that you can plug this uh this uh amateur radio amplifier into it. Um it will push it into the coil and then switch over and you hear the receive chain. Then what we had was a bunch of again relatively cheap mini circuit amplifiers giving about 60 to 80 dB of gain before we bent back into the FGPA to get the signal we needed. Um also in this box you had our gradient amplifier. So the things I was talking about earlier where you need to change a magnetic field, you need to have a very fast acting amplifier. So audio amplifiers are at roughly the sort of frequency you need to do it at, but it needs to be very clean. Uh so we used an open source project to use that. We kind of got one off the shelf and built it. Um and that lived in that box. And then the top box is the FTPA and the amplifiers and all of our clean power supplies, uh power sources. So we had a 24volt battery for our gradients and then a 12volt battery for all of our offchain. And then we split that down to all the various different voltages because electrical engineers like everything being at different voltages. So we had 12 volts and 5 volts and 3.3 volts. And then finally we've got our box at the top which is where you actually put the samples inside. Um and that's a far cage. We've sealed around the outside of it. And so you open it, you put your sample in and you scan it and then you repeat lots of times. And the more you repeat it, the better signal you get. That kind of thing I said earlier, the theoretical is proportional to the square number of repeats. We found that is the case in what we're doing, which means our noise is not from external sources are coming in. It's usually from like your amplifier being bad or um the thermal noise which is just something existing at the temperature will have a certain amount of noise. Normally you wouldn't notice it but again at nano volts you do definitely notice that. So why did we do this? Um, so this is a graphic that I stole from cancer research UK and I don't know if you know this um but for lots of cancers really early diagnosis really improves your chances of survival. So here are three different cancers and you can see on the left that's if you diagnose at the earliest stage your chance of surviving and on the right if you uh catch it late your chances of surviving are much lower. So for lots of um cancers really it's a timing question. If you can catch it early enough you can provide good treatments. I mean you can also treat people who didn't need to be treated which is another problem but um that is one of the key reasons why early diagnosis is good. If you catch it early enough you can treat it and MRI is one of the better ways of uh treating it uh detecting it. Sorry. So for here here's a bunch of common uh cancers and you can see that MRI picks them up more accurately than the existing methods and these are kind of like this isn't breast cancer normally. This is if breast cancer if you've got dense tissue. Um so it's not saying mamograms are bad or anything. This is always kind of a trade-off. But the reason why we don't use MRI isn't because it's a worse technology. It's because it's too expensive. So that brings us to what we're trying to do which is make MRIs much cheaper. Um, and our idea behind this is if we can make cheaper and more pleasant scans, and the scanner we built at the moment for 5K, you can't fit a person in, but we're hoping to scale it up and use the same sort of technology to make one that you could fit a person in. You can uh give people cheaper scans, which means you can get more data, which means you can start to collect models, which means you don't overdiagnose people. So, you can scan them early enough to work out when seeing something in someone is going to be a cancer or going to be a problem and when it isn't. And then you that means that you can scan more people because insurers will start paying for these scans which means you get more data. We need more better models. And I think the vision is just scanning everyone frequently enough that you you know what these cancers look like. Um so in a hospital at the moment MRI scan would cost about 1.5 million. I don't think you can make them for 5k but I think you can make them much cheaper than that. And that's one of the things that we kind of set out to demonstrate is you couldn't can make these things much cheaper. Um, so, uh, I, the reason why I'm doing this talk is because I'm working at a startup trying to make MRI scanners. Um, which is kind of how we got started with these things. So, if you want to get involved, email me at john@flux.cl clinic. Um, also come and chat to me afterwards if you're interested. I can explain all of this stuff in much more detail. This has been a very kind of like whistle stop tour of a lot of different things. And I thought also I would at the end give a bunch of acknowledgements to a bunch of different people. So um the gradient board was the the GPA FHDO um board. Um if you Google that you'll find it. Um and the TR switch we got from another open source project which is the AIM for II OSI project and in general there's this uh community called the OSI or open source imaging project and they've done a bunch of this stuff already. We're not we weren't like designing from scratch um but we were building on what other people had built and people there were really helpful and they gave us advice and guidance on all the various different things that we did. So shout out to them. They were really great. Um and yeah if you've got any questions I think I will be at the FQ or just grab me afterwards as well. Thanks [applause]