Stat, there’s a problem with my fMRI.

Stat, there’s a problem with my fMRI.

There are multiple versions of this news article floating around G+. I’m not sure if the popularity is due to schadenfreude or people applauding scientists for ever refining/correcting their work. Nevertheless, there are many things to correct and clarify in news about the “fMRI bug”.

It’s important to note that this is only for BOLD MRI, often called fMRI (functional MRI, which is a misused label) and this is nothing really new. Let me explain. I’ve written a lot about fMRI and it’s drawbacks before (see links below). If you aren’t familiar with what MRI is, you’ll also want to see the Medical Imaging 101 on MRI (below).

There are two main imaging tools for imaging brain activity. BOLD MR and PET imaging.

BOLD-MRI stands for Blood Oxygen Level Dependent Magnetic Resonace Imaging. Some people use BOLD-MRI and fMRI interchangeably, forgetting that there are other functional images. BOLD takes advantage of the T2* effect of deoxy vs. oxygenated hemoglobin in blood, i.e., the iron in hemoglobin has different magnetic properties depending on whether oxygen is bound or not. Positron emission tomography (PET) with F-18 labeled fluorodeoxyglucose (FDG), measures changes in metabolic activity. Putting the bug in statistical software aside for the moment, it’s important to note that neither BOLD-MRI or FDG-PET measure brain activity directly. They measure oxygenation and metabolism, respectively. Also, note that PET directly measures metabolism, which we infer brain activity. BOLD-MRI indirectly measures blood oxygenation, which we then infer brain activity. You add on top of that issues with post processing and you can see where I have little faith in BOLD-MRI.

Here are to older post that demonstrate that the errors in BOLD-MRI post processing software are not new. In the first link, a graduate student made a mistake and noticed that the resulting image looked like any previous image she acquired. The second link talks about how the size of the smoothing kernel can effect the data and it’s worth discussing because image resolution has improved and the kernel size has an effect.

When a blob is just a blob

http://goo.gl/Fx74F

Not functional fMRI

https://plus.google.com/u/0/+ChadHaney/posts/CP3dEwRWg7X

The discussion on voxels in many version of this news is misleading. A voxel is simply a volumetric pixel. Most people know what a pixel is and it’s 2D. For example, your cell phone camera might have 8 megapixel resolution, i.e., 8 million pixels from 8 million sensor elements. A voxel is an element in a 3D image (don’t be confused by 3D vision, i.e., 3D movies). Nearly all medical images are 3D. However, to save time, the third dimension, which we often call a slice, is thicker than the in-plane resolution. So a 100 micron in-plane resolution MRI scan with a slice thickness of 0.5 mm, would have a voxel size of 100 × 100 × 500 micron.

Another misnomer is “fMRI machines became available in the early ’90s”. There is no such thing as an fMRI machine. Any MRI scanner can do fMRI and fMRI means more than just BOLD-MRI as mentioned already.

In terms of a ‘replication crisis’ mentioned in the news article, they fail to mention that there is and has been a ‘funding crisis’ so scientists can’t replicate experiments even if they want to.

Medical Imaging 101 pt 3: MRI

http://goo.gl/UVbiU

fMRI – What is love?

http://goo.gl/7gxK2

Whale of a story

http://goo.gl/xcAqX

Meta-consciousness Brain

http://goo.gl/MgJiA

Cluster failure: Why fMRI inferences for spatial extent have inflated false-positive rates

Eklund A, Nichols TE, Knutsson H.

Proc Natl Acad Sci U S A. 2016 Jun 28. pii: 201602413

http://www.pnas.org/content/early/2016/06/27/1602413113.full

#fMRIblob #ScienceSunday  

h/t Kee Hinckley 

http://www.sciencealert.com/a-bug-in-fmri-software-could-invalidate-decades-of-brain-research-scientists-discover

39 thoughts on “Stat, there’s a problem with my fMRI.”

  1. Chad Haney

    I don’t have a question. I was expecting important information, but there isn’t any.

    We will have to wait for that.

  2. SharingFOCUS​, if you don’t think that there isn’t any important information here then I suggest you stop following my posts or write something more substantial yourself.

  3. Chad Haney

    I’ll decide that after watching what else comes along. You needn’t act insulted, as my definition of important is entirely subjective.

    I wasn’t trying to diminish you or something.

  4. Thanks, Chad Haney​ for sharing this important news. A 70% error rate..well, how booked up is your machine? They’re gonna be busy a long time.

  5. Ray of Sunshine​, I’ll block before your popcorn is done popping. I’m a bit surprised because this isn’t even a controversial post like vaccines or GMOs.

  6. Thanks Chad Haney. My very first thought when I saw these stories cropping up was – literally – I wonder what Chad’s take on this is?

  7. The problem with reproducibility makes me really sad, given how hard it is to get the first study funded. But we’d get such better results if we funded it along with the study.

  8. Isn’t it shitty when one single engineering flaw has the potential to annihilate so much hard work? Don’t feel bad, fMRI, PCR machines know how you feel.

  9. BOLD MRI with all of its warts is still a useful tool. It’s just annoying that many people have no clue about its limitations and talk about the results as if they are definitive.

  10. Filinto Rodrigues​, be more specific. I do a lot of work with 3D images and there are many instances where it would be a novelty but not essential.

  11. We do it all the time. How would 3D be better? Keep in mind, that a 3D image data set does not have to be visualized in 3D. On the contrary, seeing internal structures is problematic in 3D. BTW, old people aren’t the only ones who get cancer.

  12. Filinto Rodrigues in looking for things like tumors, 3D would be less effective than 2D layering. You’re looking for small changes between very thin layers of cells sometimes. A 3D image may actually have too much overwhelming information.

    If you do want to engineer something that would help find tumors and that sort of thing, they’re always looking for an algorithm to automatically scan though the test images and look for certain markers. Reducing the time it takes a human to do it. Problem is, different types of cancer looks different, acts different, etc. So as yet, there’s no one size fits all programs to help ease the eye strain of folks like Chad Haney​.

  13. Filinto Rodrigues the other issue is using contrast too often. Having nodules in my own lungs and being about the same age as Chad, I’d prefer not to subject myself to radiation as often as monitoring needs to happen.

    And yah, nodules show on just an X-ray or an mri without contrast and can be measured in reference to a corresponding mri. Btdt.

  14. Filinto Rodrigues, what do you mean by zoom-in and zoom-out? All of the visualization tools at MRI workstations allow you to zoom in and out. I’m also not sure what you mean by 3D rotation axis. 3D surface rendering is often neat but not very helpful diagnostically.

Leave a Reply