Sorry, Alan - got swamped with BIRN Ontology & Mouse BIRN AHM mtg
preparations for next week.
You are right - you & I reviewed details related to ABA image meta
data last weekend - NOT brain region level meta data.
I'd bet a lot of what Nigam lays out below - RGB LUT values and PK -
is correct.
Region Abbrev (Cols B & C):
'CNU' can be found in the Swanson (2004) XML file ("Cerebral
nuclei"). Despite the ABA atlas gray scale image plates having been
derived from the Franklin & Paxinos adult C57Bl/6 atlas, as Mihail
mentioned, the Swanson lab did some region classification for ABA,
and they lumped this at a higher level in a region they called
"Cerebral Nuclei" - though, again as Mihail points out, Striatum
immediately maps in rodent to a structure referred to as the
"Caudoputamen" (http://brancusi.usc.edu/bkms/brain/show-braing2.php?
aidi=129). Caudoputamen [very important for the PD use case] and
other structures in the base of the telencephalon are a part_of a
larger structure they've defined for ABA as "Cerebral Nuclei (CNU)
The one fly-in-the-ointment here is the phrase "...they've defined
for ABA...". This doesn't necessarily map to any of the other brain
region classification schemes/CVs used elsewhere - not a trivial
process - but not impossible - both the NN group and BIRN groups are
working on this - as is Mihail - as he mentioned). As Mihail points
out - and you can see in the XML files he distributes - it does link
into the vocabulary used in Swanson 1998/2004 for rat. It most
likely does not deterministically map to the CVs used for brain
region by GENSAT (believe that somehow derived from some combination
of NN and regions as given in the Franklin&Paxinos mouse atlas -
believe the GENSAT segmentation was performed by someone from George
Paxinos's lab who went to work with the Rockefeller GENSAT group).
SenseLab has much less brain region detail. It MAY be using the
Swanson nomenclature. Given SenseLab has just a subset of the
regions you'd find in an atlas, it's possible someone there at Yale
could fairly quickly provide a lookup table mapping their region
terms to one of the other atlases (Luis may even have done this
already in the context of some of the neuroinformatics repository
integration work he has done over the last several years).
Region color (Cols D, E, F):
All digital atlas have a color LUT for regions. These are generally
just 8-bit (only because few atlas projects have foreseen having the
expert personnel resources to manually segment > 256 regions) In the
atlases, the regions derive from very laborious manual segmentation
done in tools like AMIRA by specially trained, highly knowledgeable
neuroanatomists. The manual segmentation is performed on 2D
sections, assembled into 3D volumes, smoothed, then added to the 3D
atlas voxel data (many atlases are not actually TRUE 3D data sets -
e.g., the Paxinos atlas used at ABA - so the re-assembly, smoothing
and integration with voxels isn't required in that case).
Anyway - ABA is obviously being forward looking and using 24-bit
values for their region LUT. Besides, when using Paxinos, they are
GIVEN the region segmentation, so the manual effort is potentially
eliminated (though the only electronic version of the region
segmentation typically must be obtained through the atlas publisher -
Elsevier in this case - and generally all the regions for a given
Paxinos image (sagittal or coronal) are just lumped into a single,
bit-mapped file. This means you must take that bitmap, run
algorithms to identify the individual regions (usually based on color
- e.g., just as you see here, each region has a specified color in
the bit map). The isolated regions can then be converted to a
geometric object format (from simple point list on to quad-tree or
oct-tree) and this is then stored separately in a RDBMS. This is
EXACTLY what the SMART Atlas project in BIRN (from Maryann Martone's
NCMIR group at UCSD - source of CCDB, too) has done. This way, each
individual region is defined as a geometry in a specified coordinate
space AND - most importantly - can then be used to support SPATIAL
queries on the atlas (e.g., "Show me all the defined brain regions
that lie within this shape I just drew on image X that you've
registered to your atlas coordinate space).
Other Region numbers:
A) Cols G & H
I would guess these are BOTH PKs of some sort as they both contain
rather small and unique integers. Given column G is listed in order,
I'd guess that is the ABA internal PK for that region. The other ID
is probably a cross-reference to another brain region classification
scheme. A search of the various atlas classifications on the
Mihail's BAMS site doesn't appear to provide such equivalent IDs.
I've searched in NN, but those IDs don't correspond either (e.g., the
Col H. for "Thalamus" - 351 - does not correspond to the NN ID for
"Thalamus" - 283).
One interesting note - if you sort the spreadsheet on Col H - you
will find the rows are ordered in nearly perfect alphabetical order
by brain region abbreviation. This indicates to me these Col H
values are likely to relate to IDs created for these regions by
Mihail/Swanson when they did this classification work for ABA. There
are no integer PKs given in the BAMS XML files from the Swanson lab
that match these numbers, so only Mihail can vet this hunch. I'd
guess they are expecting to use the "brain region abbreviation" as
their immutable, unique link.
B) Col I:
Typically, the whole purpose of registering brain image stacks into
the coordinate space of a digital atlas (such as has been done for
the 20,000+ ABA image stacks for the individual, gene-specific,
GenePaint-ed brains) is so the expert segmented regions from the
digital atlas can be used to drive QUANTITATIVE ANALYSIS of the
registered image stacks in a consistent and comparable manner. Being
able to visualize the atlas regions overlayed onto individual images
from a given registered stack is useful for making qualitative
observations - or as a pedagogical aid - BUT it can't drive automated
analysis unless:
- the atlas comes with a coordinate system and the segmented brain
regions have been deterministically mapped into that space
- the image stacks have been registered to the same coordinate
space.
In the case of the ABA, the atlas is the Franklin&Paxinos 2001 adult
C57Bl/6 brain atlas and the coordinate space is their interpretation
of stereotaxic coordinate mapping. The registration process has some
error (for ABA, I believe that is 300 microns [probably different for
in-plane registration - i.e., 2D to 2D alignment - vs. the third
dimension between images in a given dissection axis (i.e., coronal or
saggital for ABA)]).
SO - for Cols J,K,L, they likely refer to the location of the brain
regions in the coordinate space. In a true 3D atlas, all you'd have
to do is give 3D geometric definition of the region shape (e.g., as
an oct-tree), and give the location of the centroid for that shape in
the coordinate system. Since the F&P atlas is NOT a true 3D atlas
but rather a series of 2D images, you don't have a 3D geometric
definition of the region. With that in mind, the way to specify
WHERE in the atlas a given region lies is to give the FIRST and LAST
atlas image plate the extents of that region lie in when viewed along
a specific dissection axis. For very convoluted structures such as
the hippocampal formation, this can be a bit problematic to use
computationally, but it's usually sufficient for the types of tasks a
2D atlas can support.
In terms of what this set up can support, it is likely one of the
things they are looking to do is to support users segmenting the gene-
specific images (GenePaint-ed images) then comparing those gene
specific segmentations to the brain segmentations. Given the limits
of the 2D realm, you can't really do true 3D volumetric
intersection. However, what you can do is determine to what extent
regions-of-interest (ROI) created by users to identify expression
patterns on a given GenePainted image overlap with the 2D sections
through the brain regions that appear on the corresponding,
registered Paxinos plate. You would essentially try to estimate an
intersection of the user drawn 2D ROIs with the 2D atlas region
shapes for each region that appears on that Paxinos plate. When
sorting those numbers by atlas brain region, you'd then go through
ALL the atlas images that contain a given region, and add up the
total intersection with the user drawn ROIs across all the images
from BRAIN X that user drew ROIs on. This would be normalized to the
total approx. pixel volume for that brain region across all the atlas
plates where it appears - resulting in an APPROXIMATE volume ratio of
a given gene stained for in BRAIN X with a given region defined in
the atlas. The large number - Col I - appears to scale with the
approximate size of the region - e.g., "Olfactory areas" is quite
large (838206), whereas one of the smaller regions included within
OLF is much smaller (Nucleus of the lateral olfactory tract [NLOT] -
7407) - so I would guess this column represents that altas-defined
approx. region volume (i.e., sum of all the 2D areas defined for that
brain region across all the F&P atlas images).
B) Cols J, K, L:
It's quite likely ABA calculated an approx. 3D location value for a
region probably truncated based on the existing locations of the
Paxinos plates within the stereotaxic coordinate space. Those
coordinates would be specified either with:
* a 2D coordinate location within a F&P atlas image plate (either
as unitless PIXELS or as stereotaxically-defined MICRONS) + a unique
ID for that F&P atlas image plate derived from a specific dissection
plane axis (e.g., F&P Coronal plate 23)
* a 3D coordinate location that somehow places some morphological
property of the region in the stereotaxic coordinate space - e.g.
front-upper-left point for the approx. 3D bounding box for that
region, centroid for the approx. 3D bounding box for that region, etc.
That's all I have time for now. Must get back to meeting prep. It's
possible reading the ABA Nature paper from January would get you a
more specific answer - or - better yet - just drop an email to the
guy you spoke with at ABA.
Hope that helps.
Cheers,
Bill
On Mar 2, 2007, at 5:15 PM, Alan Ruttenberg wrote:
Don't recall doing this, though it's certainly possible that I've
forgotten.
Just to clarify, each of these lines is for a brain region, not for
an image.
If you want to do this later this evening with me, give me a call
at home after about 10.
-Alan
On Mar 2, 2007, at 5:10 PM, William Bug wrote:
Alan,
Didn't you and I review this already at the ABA site.
All one would need to do is bring up one of these images at the
ABA site, go through the "noodling" we did, and look at the
corresponding entries in the spreadsheet to match up a "meaning"
to each column (probably nearly all those columns).
Cheers,
Bill
On Mar 2, 2007, at 2:15 PM, Nigam Shah wrote:
BTW, if someone has a theory of what the other number in
ontology.xls are, I'm all ears.
Okay, pure guesses:
Line 4 = Cerebral
cortex,CTX,CH,176,255,184,3, 85,4141526,61.647,29.999,33.711
176,255,184 seem like RGB values (they all range from 2 to 255) for
that region in the image.
3 is a serial number or internal id.
85 - no clue
4141526 - no clue
61.647,29.999,33.711 seem like 3D voxel coordinates.
--Nigam.
Bill Bug
Senior Research Analyst/Ontological Engineer
Laboratory for Bioimaging & Anatomical Informatics
www.neuroterrain.org
Department of Neurobiology & Anatomy
Drexel University College of Medicine
2900 Queen Lane
Philadelphia, PA 19129
215 991 8430 (ph)
610 457 0443 (mobile)
215 843 9367 (fax)
Please Note: I now have a new email - [EMAIL PROTECTED]
Bill Bug
Senior Research Analyst/Ontological Engineer
Laboratory for Bioimaging & Anatomical Informatics
www.neuroterrain.org
Department of Neurobiology & Anatomy
Drexel University College of Medicine
2900 Queen Lane
Philadelphia, PA 19129
215 991 8430 (ph)
610 457 0443 (mobile)
215 843 9367 (fax)
Please Note: I now have a new email - [EMAIL PROTECTED]
