Proposal for MR data classification

[gsfr]

The purpose of this classification proposal is to capture how users think about their data and how they would want to search for it, but not to be a ground truth, MR physics-based classification that could be used by processing algorithms.

MR data will be classified according to the following aspects, which are all multi-select.

• Contrast: B0, B1, T1, T2, T2*, PD, MT, ASL, Perfusion, Diffusion, Spectroscopy, Susceptibility, Velocity, Fingerprinting
• Intent: Localizer, Shim, Calibration, Fieldmap, Structural, Functional, Non-Image
• Features: Quantitative, Multi-Shell, Multi-Echo, Multi-Flip, Multi-Band, Steady-State, 3D, Compressed-Sensing, Eddy-Current-Corrected, Fieldmap-Corrected, Gradient-Unwarped, Motion-Corrected, Physio-Corrected, Derived, In-Plane

---

Beyond Neuro MR...

• Specimen: Phantom, Human, Animal, [or list of animals]
• Anatomy: Head, Neck, Upper Extremities, Chest, Abdomen, Spine, Pelvis, Lower Extremities, Other

---

Next steps:

• get input from more users
• define mappings for all BIDS types
• consider renaming measurements DB key to classification

[chrisgorgo]

Mapping of the basic BIDS data types:

• _T1w : Contrast == T1, Intent == Structural
• _T2w : Contrast == T2, Intent == Structural
• _T1map : Contrast == T1, Intent == Structural, Features == Quantitative (however we are considering just storing the INV1 and INV2 from the MP2RAGE sequence and leaving the reconstruction to the user)
• _T2map : Contrast == T2, Intent == Structural, Features == Quantitative
• _bold : Contrast == T2*, Intent == Functional
• _dwi : Contrast == Diffusion, Intent == Structural

Comments/Questions:

• There are more structural scan types in BIDS, but I only focused on the most common ones.
• Discussion about how to model multi-echo data in BIDS is still ongoing.
• In your nomenclature is multi flip angle considered multi-echo?
• How do you model sequences like MP2RAGE with multiple inversion times?
• How do you model fieldmaps? In BIDS there are 4 different ways data useful for generating fieldmaps can be stored: https://docs.google.com/document/d/1HFUkAEE-pB-angVcYe6pf_-fVf4sCpOHKesUvfb8Grc/edit#heading=h.fcegd01wpsf8
• I would suggest leaving "Corrections" out of this spec and focus on raw data coming out of the scanner. Preprocessing is another can of worms.

[gsfr]

Thanks, @chrisfilo. I'm not an expert, but would kindly ask @rfdougherty for his thoughts on your comments.

Please note that the purpose of this classification proposal is to capture how users think about their data and how they would want to search for it, but not to be a ground truth, MR physics-based classification that could be used by processing algorithms. I'll also update the original post with this language.

[rfdougherty]

I think the mapping looks reasonable. As for the comments/questions, I first have a general comment (maybe 'rant' is a better word). I find it confusing to conflate the raw files used to compute a quantitative metric (T1, T2, fieldmap, etc.) with that metric. E.g., calling the multi-inversion scans or mulit-flip scans '_T1map'. I think just storing the raw images as generic structurals is better, reserving 'XXmap' for the quantitative map that is computed from them.

• I would keep multi-flip distinct from multi-echo. They mean very different things in terms of the pulse sequences and are also generally used differently by end-users. (multi-flip for quantifying T1, multi-echo for qT2 or BOLD artifact removal).
• As noted above, I'd just store multi-inversion, multi-sflip, etc. as structurals, with the relevant parameter available so that they can be easily used to compute the quantitative map. It might be useful to be able to mark a set of images that form a useful 'group'. E.g., the multi-inversion scans are grouped together so a processor can easily grab them to compute a T1map.
• Same for fieldmaps-- I'd reserve 'fieldmap' for the actual fieldmap (I think that's your 3rd option), and just 'goup' the raw images useful for computing a fieldmap.

[gsfr]

Thanks, @rfdougherty. I added multi-flip to the features above.

Do we also need a 'fieldmap' feature, akin to quantitative? Not to be confused with fieldmap-corrected.

[gsfr]

The answer is no. Fieldmaps would be represented as "B0, quantitative" or "B1, quantitative".

[mattmiddione]

Compressed Sensing, Fingerprinting, Localizers, High Order Shim, Calibration Scans, Anatomy, and Perfusion vs. ASL

[gsfr]

The existence of intent edge cases, such as functional localizers has been pointed out. In response, I'm removing the single-select restriction on intent and hence simplifying all classification aspects to multi-select.

[gsfr]

Merged features and corrections, per discussion with @lmperry.

[gsfr]

@lmperry Per today's conversation, should we add a Derived feature here?

[gsfr]

Added Derived feature above.

@nagem Please incorporate accordingly.

[gsfr]

Added Non-Image intent.

cc @ryansanford, @nagem, @lmperry

[ryansanford]

FYI @dpuccetti DICOM Viewer should not launch for Non-Image intent on dicom types

[gsfr]

@nagem Added In-Plane feature.

[gsfr]

@nagem Added Fieldmap intent.

[ehlertjd]

I think Phase fits in as a feature -- used in Susceptibility Weighted Imaging.