Why do MRIs still take so long?

June 23, 2009 11:34 AM

Why do MRI's take so long? Just went in for an MRI of my knee (fingers crossed), and I noticed that MRI machines have gotten significantly smaller (this one was the size of a refrigerator). But the scanning process still took more than 30 minutes.

So while I was sitting there, mind wandering, I started thinking - what is it about MRI's that take so long to complete?

Difficulty Level: I'm a computer guy through-and-through, I know the basics of how an MRI works, but I know nothing about the actual mechanics/engineering.

In the 20+ years since MRI's first became available, computers have increased in performance more than tenfold. Perhaps not as drastically, we have also seen huge improvements in robotics, image sensors, precision manufacturing, etc.

I'm assuming MRI scanners are significantly cheaper than they used to be - especially because you can now buy small units for individual body parts, rather than full-body scanners (in my case, the ONI OrthOne).

So what's the deal? Has performance (time) been held more or less constant while manufacturers focus on performance (size/image quality/price)? Or is there some kind of inherent barrier to speeding up MRI's?

Really curious as to any insights as to where MRI/scanning technology is headed.

posted by swngnmonk to technology (6 comments total)

(Take everything here with a grain of salt; I'm a pure mathematician with an amateur interest in this stuff, not a professional imager.)

As it happens I've recently been learning a lot about why it's hard to do fast MRI. To oversimplify the science (and no doubt to mangle it to a certain extent) you're up against a physics problem, not a technology problem. You're using a pulse to reach into the body and "twang" the magnetic fields of some atoms inside there, then recording what happens as they relax back to their unexcited state. How long that takes is a property of the atoms, not your MRI scanner or the software analyzing the images.

Given that hard limit, the progress in fast MRI isn't about doing a single scan faster; it's about getting a reasonably sharp image with fewer scans, by exploiting prior knowledge about the expected nature of the image you're trying to reconstruct. Chuck Mistretta's lab here at Wisconsin is doing some really interesting work on this, getting speedups in the 10x to 1000x range for certain types of imaging problems; here's a press release which should give you enough keywords to search on if you want to look at some more technical material.

By the way, the tradeoff, as I understand it, is always image quality vs. time -- so any new algorithms that allow the scans we currently do to run faster will also get much higher-quality images out of a 30-minute scan; so I think at least those two goals, of the four you mention, aren't in competition with each other. As for size and price, I'm not sure; my sense is that the action is on the image-analysis side more than the hardware side, so I don't think you should think of the new MR as involving bigger, more expensive, or more powerful scanners.
posted by escabeche at 11:49 AM on June 23 [1 favorite has favorites]

Did you read the wikipedia page? It sounds like...it just takes a long time, physically. Specifically the section on the pulse sequence:

The first part of the pulse sequence, SS, achieves Slice Selection. A shaped pulse (shown here with a sinc modulation) causes a 90° (π/2 radian) nutation of longitudinal nuclear magnetization within a slab, or slice, creating transverse magnetization. The second part of the pulse sequence, PE, imparts a phase shift upon the slice-selected nuclear magnetization, varying with its location in the Y direction. The third part of the pulse sequence, another Slice Selection (of the same slice) uses another shaped pulse to cause a 180° (π radian) rotation of transverse nuclear magnetization within the slice. This transverse magnetisation refocuses to form a spin echo at a time TE. During the spin echo, a frequency-encoding (FE) or readout gradient is applied, making the resonant frequency of the nuclear magnetization vary with its location in the X direction. The signal is sampled nFE times by the ADC during this period, as represented by the vertical lines. Typically nFE of between 128 and 512 samples are taken.
The longitudinal relaxation is then allowed to recover somewhat and after a time TR the whole sequence is repeated nPE times, but with the phase-encoding gradient incremented (indicated by the horizontal hatching in the green gradient block). Typically nPE of between 128 and 512 repetitions are made.

Typically TE is between 5 ms and 100 ms, while TR is between 100 ms and 2000 ms.

The diagram in wikipedia seems to conflict slightly with the text, its unclear if TR (or if the 100-2000ms of TR is just recovery time and the whole cycle is long. includes TE, lets be conservative and say it does, and the max for each repetition is 2 seconds, and you are doing a max of 512 repetitions.

If I read this right that would be up to 17 minutes for one slice! Never mind how many slices they do. But the limitations are in the physics of the how long it takes the spun particles to revert.
posted by jeb at 11:59 AM on June 23

I've always been told that MRI is just like NMR, in which case my first guess is something along the lines of: more scans = higher signal-to-noise ratio (better and sharper image). But I'm totally guessing.
posted by KateHasQuestions at 12:53 PM on June 23

Given my somewhat limited knowledge for using them for research, the above people are right that it is a physical limitation of how long it takes the atoms to reset, but it is also a bodily limitation, because basically the faster you go, the more energy you end up putting into the system (your body) and there is a limit of how much energy your body can take (you can actually start to heat up the body).

So there are some safety issues, plus scanners are tending to become stronger (so we can get better pictures), but the increased strength means you can't necessarily speed it up, because of the limitation I mentioned above.
posted by katers890 at 1:08 PM on June 23

I used to know some grad students who worked on medical imaging technology and they were focused on what escabeche is talking about: use less imaging data but better algorithms (exploiting what you already know, probabilistic techniques, etc.) So..."I heard this from a dude in a lab years ago" isn't exactly data, but it anecdotally supports escabeches point.

Sidenote: One of the projects a guy was working on wasn't based on MRI, it was based on taking intravenous ultrasound video and reconstructing animated models of the functioning vascular system. Think about this: there is not enough data: the probe is moving through the artery (I think by hand, so the speed is probably all whack) and getting a 2D slice of the artery at an instantaneous time. Somehow, this guy was reconstructing the 3D structure and motion of the blood vessels being examined, basically by using heuristics, probabilistic approaches, and such. It was awesome! You could see him load these blurry IVUS files into his software and then it spits out a 3D pulsating coronary artery model. Plus, since he was recovering the motion data, it was (supposed) to help doctors better understand where arterial plaque, aneurysms or other arterial degradation was occuring because it changed the movement of the arterial walls. This was in a lab like years ago so its probably in the field by now. He was a computer science guy, so some of the 'this will help doctors' was I bet speculative. Oh, that reminds me: him and some other guys later extended this technique to reconstruct fluid flows in the artery and then loaded the data and viz software into a CAVE, and you could don the polarizer glasses and move through the artery, and then use the remote control to shoot vectors down the artery that would follow the fluid flow at that point. It was soooo cool! I have no idea if this is useful in any way beyond like...getting dataviz grants though.
posted by jeb at 1:19 PM on June 23 [1 favorite has favorites]

If it was the size of a refrigerator, then it was probably one of the lower power machines attached to a fairly simple controller.

The detailed anatomical scans can take a while because you want fairly small slices to obtain a high resolution. They may also have taken multiple scans, as the frequencies you use to image bone are different than those you would use for soft tissue. Just my guesses though. I've had some grad. school experience with functional MRI, but I'm by no means experienced with the medical end of it.
posted by Avelwood at 2:29 PM on June 23