CT Data – Finally talking to a radiologist

So I finally sat down with one of our radiologists to pick his brain about what is going on with this data and why do I get the data that I get.  I’d mention him here because he was tremendously helpful, but I don’t want anything I say to make more work for anyone.

What did I learn?

1) Variable slice thickness is most likely a myth.  

The reason I said this is because he told me that the data is likely acquired using 0.625 or 1.25 millimeter (mm) slice thicknesses (really thin cheese).  That data is then usually outputted to whatever slice thickness the radiologist needs to read.  That can usually go anywhere from 5mm to 10mm.  Changing this slice thickness allows radiologists, who read a whole lot of scans per day, only have to read 30-60 slices vs. the 400 or so that the raw data contain.  

Now, back to the variable slice thickness.  From what I can tell, this happens because there isn’t a lot of action (disease/injury) going on in some parts of the brain, so they don’t need to review those places.  Also, there is a bit of liability where “you can’t miss what you can’t see”; you may have seen something on the thinner sliced scans.  So overall, I get it – you don’t want to spend your time reading 1000 slices at a time and that’s totally reasonable.

That said, we discussed the fact that the raw data should be on their PAC system or a parallel server of storage.  This data can usually be retrieved by asking one of the techs for it.  Also, you could theoretically ask for all the data in 5mm slice thicknesses; we’re considering standardizing this in our protocol of our recently-funded upcoming Phase III clinical trial: MISTIE III.  (Side note: why do we reference phases of clinical trials by Roman numerals?).

2) Gantry tilt.  So this is really to protect the eyes during conventional CT scanning.  I mean who wants to come out of a stroke to only hear “You’re fine…but you may have some cataracts”.  There has been some research on trying to correct it using some projections (http://www.ncbi.nlm.nih.gov/pubmed/17282566, as well as a few others but this has some good pictures).

3) Acquisition.  So my analogy with the cheese is still valid for conventional scanning, but helical or spiral CT scanning is more common.  Think spiral ham.  The scan is taken while the X-ray tube spins around the table the patient is on.  There are reasons to use this method over conventional scanning, but I’m not going to mention them here.  

4) Registration – so the discussion of registration was somewhat light.  We discussed the use of SPM and FSL in image registration, but the discussion was slighted to more MRI which don’t have some of the same problems.  We did discuss registering MRI to CT, but this is commonly done using BrainLab, an intra-operative program that helps image guidance, and is done by using external landmarks on the head.  While this is a method that may prove worthwhile, not all CT have the landmarks – usually only the operative scans do.

So overall there’s a lot out there left to do, but it’s good to hear that there is some rhyme and reason to getting data in this way.  Next steps – make friends with some techs to find out how the best way to get a multi-center trial to tell it’s coordinators that they need to request these.

CT Stroke Data – what was I thinking?

So I’ve been working with some CT data (brain) after working a few years in MRI.  I’ve found this paper http://bit.ly/13idoBH that talks about the problems of acquiring CT and the DICOM (Digital Imaging and Communications in Medicine) format.  I’ll discuss some of the things I’ve encountered trying to get these images into formats that my MRI-based tools can understand.

For those that don’t know about image acquisition, imagine you’ve got a block of cheese and you’re at the deli.  The slicer is the CT scanner,  the cheese – your brain.  Each of the slices is one DICOM image.  Nice cheesy DICOMs.  I want a sandwich.

To those who do imaging, here’s a couple of things I think:

1) MRI’s are spoiled.  They are usually acquired with a lot of specifics and usually are well documented.  They also have a lot of tools to work with them.  I didn’t know how good I had it for a while.

2) Variable slice thickness should be banned.  This refers to you take 5mm slice scans at the bottom of the brain and change it to 10mm later in the scan (and maybe change it back!).  Stop this madness.  That’s like asking the deli guy (or gal) “Hey I want a 1/4 pound thick and then 1/2 pound thin slice and then 1/4 thick on top of that”.  Now I understand that viewing things in finer detail is necessary/desirable and that CTs have low-dose radiation so shorter scanning times/fewer images is better.  

That being said – a lot of patients (and CT is usually a diagnostic population – so yes, patients) are getting upwards of 10 scans.  You’re dosing them.  I’m not glad, but if you’re going to do it and you need that finer resolution, just set the whole scan to it.  I don’t need a fine section of their nose, but stop playing with the knobs and giving me a weird block of cheese.  

Now what to do?

“Duplicate the slices so that you fill in the gaps and have the finer resolution” – have you seen what that looks like in 3 dimensions.  No – you haven’t, because it doens’t look correct.

“Interpolate the data and make it a uniform thickness” – I’ll just average your eye and frontal cortex – looks nice.  This isn’t a bad idea – but I think staying with rawer data is better.

“Well just take out the duplicate slices” – throwing out data is not what I’m into. Not the worst option though.

“Don’t use that scan” – get out, no seriously, out.

3) Gantry tilt.  I didn’t know I was supposed to hate gantry tilt, partially because I didn’t know what that was.  The paper above describes that they do this to help shield vital organs (like your eyes – not your brain) from radiation.  Pretty much you shoot X-rays in a different way that isn’t perpendicular/orthogonal to your head.  

Now if you’re saying they just acquire the scan at  “a funny angle”, I’d say to you “It’s behind you Tyrone. Whenever you reverse, things come from behind you.” http://www.imdb.com/title/tt0208092/quotes.  

Anyway – looking at a scan that was acquired with a bad gantry tilt pretty much looks like one of the cone heads walked into the scanner.  The sagittal view of an axially acquired scan looks like you stretched the posterior of my head with a plunger.

Ok – what to do – there are some simple transforms to cram your head back to normal: http://www.mathworks.com/matlabcentral/fileexchange/28141-gantry-detector-tilt-correction.  Thank you for such scripts.  

Are they perfect?  What a terrible leading question – of course not.  Images are acquired in a FOV (field of vision) and if you rotate the head, you may push some data (like your frontal cortex) out of it.  There are some shifting you can do to try to combat this.

4) DICOM Header information is a plethora of information…and not de-identified enough.  Pretty much they are “metadata” (https://en.wikipedia.org/wiki/Metadata), or information about how this data was acquired. DICOM headers have very vital information, such as the actual degree of gantry tilt described above, the date/time the scan was acquired, what type of scan it is.  It has information specific to each slice (of brain/cheese), which can be read with some well-established tools (DCMTK is good: http://dicom.offis.de/dcmtk.php.en).  

But with all that information, different hospitals like to cram a bunch of identifiable markers in there, and not always in standard places.  Patient’s names, doctor’s names, birthdays, OTHER patient names, peoples phone numbers, their favorite colors, hobbies, etc.  This can become troublesome for hungry hungry HIPAA.  

5) There is a light at the end of the tunnel.  The data is rich, very common, and it’s cheaper to acquire than MRI.  “Hey I went the ICU” – CT scan, car accident – CT scan, brother hits you with a rock – CT scan.  


Hopefully I’ll have another post about how we can hurdle these challenges.  Then maybe that sandwich.