We use inductive detection. This is something you're all familiar with as a child, where you take a magnet and move it through a loop of wire, and you generate a voltage. In this case, the moving magnet actually comes from the nuclear polarization of the water protons typically in your body, and, in fact, what Richard Ernst calls "the powers of evil" has to do with the fact that nuclear magnetic moments are very, very small.
If you're interested in making images of this quality, over the span of a few seconds or minutes, it means you need to make this multiplicative term B, the magnetic field, very, very high. That means that all clinical scanners operate, in the tesla range, typically around 3 tesla. Knowing that, what sort of images do you think we'd be able to make at our field strength, roughly 500 times lower magnetic field, which is a calculated SNR of around 10,000 times lower? Well, you'd probably guess that we couldn't make very good images, and, in fact, you'd be right.
Up until a few years ago, these were the kind of images we were making in our scanner. This is, in fact, a human head, if you can believe it. It's a single slice, took about an hour to acquire, and nobody was very interested in this at all.
If this is all we had, I wouldn't be here today, so let me tell you how we solved these problems.
Really, how do you solve a hard problem? What we've been working on is a suite of technology, half of it based in physics, half of it based in the availability of inexpensive compute. The physics applications are really about improving the signal strength, or the signal-to-noise, coming out of the body and into our detectors, and then the compute side is really about reducing the noise or getting more information from the data we have, or fixing it in post, as some people in this audience might call it.
Let's start really at the beginning, our acquisition strategy. The way you do NMR, or at least the way we do NMR at, remember, very, very low magnetic fields with very, very low signals, is we take our magnetic field. We turn it on. In red is that very small nuclear polarization I talked about. We apply a resonant radiofrequency pulse. We tip the magnetization into the transverse plane, and then we apply a series of coherent radiofrequency pulses to drive that magnetization back and forth very, very rapidly.
Then, again analogous with this inductive detection approach, we detect our signal, but not using a giant hand and a magnet moving, but instead using a 3D printed coil, in this case around the head of my former colleague, Chris LaPierre, to detect this very, very small, but with a very high data rate signal. We call this Balanced Steady-state Free Precession. That's a bunch of words. What it really means is that we now have an approach to very rapidly sample this, although very, very small, signal coming from the head.
What this has allowed us to do is to make images like this. In six minutes, we can make a full 3D dataset, roughly 2.5 millimeter in plane resolution, 15 slices. Just remember, this is the same machine, okay? The difference between these images has to do with the way we interrogate the nuclear spins, the fundamental property of the body of water, in this case, and the way we sample it. That's pretty nice. Having a high data rate actually allows us to now build up even higher quality images by averaging, and those are some images shown here, but there are other approaches, and this is where we start really talking about compute.
Pattern matching is an interesting approach people are very familiar with in the machine learning world, but we all know about this from basic physics. As an example, think of curve fitting, which you could think of as pattern matching. Curve fitting, you have some noisy data, shown as open circles here. You have some model for the way that data depends on some property, say time, so you take your functional form. You fit that function to the data, and you extract not only the magnitude of the effect, but also additional information, in this case a time constant of some NMR CPMG data.
The MRI equivalent of pattern matching is known as magnetic resonance fingerprinting. In contrast to what we did above, where we add up all of these very noisy images to make a higher quality image, in this case, we don't average actually. We just acquire the raw data. You see the data coming in to the lower left. These are very, very noisy, highly under-sampled images that normally you would sum together. The interesting thing we do here is we sort of dither the acquisition parameters a little bit. In the upper left, we show exactly how much we tip the magnetization, and in the upper right, we vary a little bit about the time in between individual acquisitions.