Would You Redo One MRI Per Day If It Saved You 10 Minutes? How about 1000?

Acceleration limits of deep learning assisted scans

January 08, 2023

Good morning!

It's time for your next dose of AI news in Radiology. I'll be your best friend for the remainder of this email; the Calvin to your Hobbes.

Today's agenda:

  • Cost/benefit for moving fast in MRI scans

  • More deep learning image reconstruction

  • Rad-Memes

I feel the need! The need for...deep learning–based reconstruction?

Need for Speed

I'm sure you've heard it before — "Do you want something done fast or do you want it done right?"

How about...mostly right?

2-D Imagines of brain from were analyzed on the impact of increasing accelerations 2x all the way up to 100X during an MRI. So, a 50 minute scan could take as little as 25 minutes. As you can imagine, this isn't a free lunch and the costs come in the form of image clarity.

This is where deep learning reconstruction comes in. By feeding a model with nearly 6000 brain MR images, deep learning can patch in the more predictable parts of the image allowing for the scanner to move faster. Think of it as the autocomplete for MR. This particular study tried to zero in on the limits of that speed vs image usability.

Radiologists found that 100% of the images were useful at 4X the acceleration. But the real kicker was 94% of the images were acceptable at 14X the acceleration. Quite the speed!

I have to wonder would it be worth it to have 14x speed in acquisitions if 6% of cases needed to be re-scanned.

Some quick analytics:

A large inpatient hospital could do:

  • ~1700 brain MRIs in a year

  • 6% of those scans are not usable

    • that is 102 scans that would need to be repeated

Would 14x increased MR speed justify having those repeats?

Something is happening in deep learning image reconstruction

As I was writing this, another study dropped on the topic of deep leaning reconstruction (DLR). Basically, DLR is the KFC double down of Radiology AI right now--double the meat.

Never ate one but kind of regret that now.

Hospital for Special Surgery in New York City looked at MRIs of the lumbar and was able to cut acquisition time by 54%. The researcher's technique will speed up imaging acquisition without giving up accuracy.

So what?

DLR has a very clear ROI for hospitals. It would be easy to explain to a CFO

  • This exam used to take 60 minutes and now it takes 30

  • We have a third next available appointment of 20 days that that exam (lots of demand)

  • We can be twice as productive without having to buy a new machine

This will be a big win once it's ready. If interested, I'll take a deeper dive on DLR and how we get from the research phase to implementation and what the biggest risks are. Just let me know.

Now some Rad Memes for making it this far!

That’s all for today!

Thanks for reading and I'll see you next week for more With and Without AI.

Mitch

If you liked this piece please subscribe below: