Source
Structure–plate assignments are taken verbatim from the authors’ published
Index of structures (structure, abbreviation, plate range), not from OCR of the plates:
Radtke-Schuller S, Schuller G, Angenstein F, Grosser OS, Goldschmidt J, Budinger E (2016).
Brain atlas of the Mongolian gerbil (Meriones unguiculatus ) in CT/MRI-aided stereotaxic
coordinates. Brain Struct Funct 221(Suppl 1):1–272.
doi:10.1007/s00429-016-1259-0 · open access.
The three plates of a level
The atlas prints every level three times: the Nissl -stained section, the
Gallyas myelin-stained section from an adjacent section beside it, and the
labeled drawing traced over them. All three are here, and Labeled /
Nissl / Myelin above the plate switches between them.
A fourth button, MRI , shows the atlas’s own reference MRI — the
imaging half of the “CT/MRI-aided” in its title. It is already in the
atlas’s coordinates, one slice per plate, so everything drawn over a plate is in
the right place on it too. Two things it is not: it is the whole head rather than a
cut section, so there is skull, scalp and eye in the picture; and its voxels are
117 µm across against the drawing’s 18, so it is soft wherever the
drawing is sharp. Unlike the other three it is fetched rather than carried inside the
page, so this file has it only when it sits beside a data folder.
They are registered to each other by construction, not by fitting. Every one of the
186 pages carries the atlas’s own printed ML/DV coordinate box, and each was
cropped to its own box, so all three land in the same coordinate space — the same
space the 6,336 label positions are recorded in. That is what makes the switch worth
having rather than decorative: the grid, the scale bar, the measure tool, the
coordinate lookup and the outlined structure are all still in the right place on a
section with no labels printed on it, and hovering where the drawing prints an
abbreviation still names it. Checked by re-deriving the labeled plates the same way
and landing on the ones already in this file, pixel for pixel, on 61 of 62 (the
remaining one is out by a single pixel).
Gray takes the color out of the drawing so the three read alike; the histology
was published in gray already. Contrast stretches whatever is showing, and both
carry into the PNG export. The 3-D view is built from whichever source is selected
there — the Nissl unless you say otherwise — and stretches it in the
renderer instead, with Floor , Ceiling and Gamma
windowing the tissue channel before Density decides how opaque what survives is
drawn.
The myelin section is an adjacent section, not the same one: the two stains
cannot be applied to a single slice. It is aligned as the authors published it.
Coordinates
Plate AP coordinates were read from all 62 plates and match the printed series exactly
(bregma +7.80 mm at plate 1, −0.35 mm per plate; lambda = bregma + 4.45;
interaural = bregma + 7.25; occipital crest = bregma + 9.95). Sections are coronal, at
350 µm intervals.
ML and DV are read off each plate’s own printed coordinate box (ML ±8 mm, DV +1 to
−10 mm), which the frame-relative cropping makes common to all 62 plates; midline
structures land within 0.2 mm of ML 0 and bilateral pairs come out symmetric. A
coordinate given for a structure is the median of its printed labels , which sit near —
but not exactly at — its centroid.
Gross divisions
The atlas names 723 structures and no containers for them — there is no
“hippocampus” in its index, only CA1, CA2, CA3, DG and their layers.
Twenty-one divisions are added here and sit above the structures in the list.
Pick one and it behaves like a structure: outlined on the plate in teal
rather than blue, shown on every plate it is on, plotted in the projection, drawn
in 3‑D as its members’ meshes, and carried by a link.
A division has no geometry of its own . Its outline is its members’
outlines with the boundaries they share dropped, its area the sum of theirs, its
coordinate the median of their printed labels — so it can never show a line the
atlas does not draw. List them narrows the structure list to one division’s
members, which is what the CSV and Labels exports then write; a structure’s own
card says which divisions it is in.
Divisions overlap on purpose : the brainstem is the midbrain, pons and medulla
together, the olfactory bulb sits inside the olfactory areas, and the four cortical
lobes inside the cerebral cortex — so two divisions’ areas do not add.
Which structures make up a division is a judgement made here, not something the atlas
publishes; every member of every one of them is written out in
data/gerbil_atlas_groups.csv, and Methods says how each was drawn.
Nine structures the atlas draws under their parts’ names on some of the
plates the index lists them on: the claustrum is Cl on plates 12–15
and 27, and DCl and VCl — never Cl —
on 16–26. The whole’s card says which parts it is drawn as and where, a
part’s card says whose part it is, and on such a plate the whole is not
“not located”: the plate says the atlas draws it as its parts there. The
table is data/gerbil_atlas_parts.csv; it is a reading of what the
plates print, not a judgement, and it is admitted only where the published index
puts the whole.
The section as a map
Color regions over the plate fills every region on it, and no two regions
that touch are given the same color — the section colored the way a map of
countries is. It shows the whole partition at once: where one structure stops and the
next starts, how much of the section each holds, which of them are neighbors. On the
Nissl , the myelin and the MRI , which carry no printed lines
at all, it is the only view that shows any of that. Wash sets how strongly the
colors are laid on; the plate is underneath them at every setting, and both ride into
the PNG and the SVG.
Touching is read off the extents themselves — two regions share the vertices of
the boundary between them — and a gap thinner than 0.05 mm counts as touching
too, because a lamina a pixel or two wide is not a boundary the eye can find, and one
color on both sides of it would read as one region.
A color means nothing beyond “not my neighbor”. It is not a
system, a division or a value: two regions of the same color on one plate have
nothing in common except that they do not touch. What a color does do is stay put
— each region asks for the same color on every plate it is drawn on and gets it
wherever a neighbor has not already taken it, which is about half the time, so
stepping through the levels is not a kaleidoscope and the same plate is the same
picture every time it is opened.
Nothing is drawn that the atlas does not draw. The fill is the region extents, which
are cut from the atlas’s own lines, and it carries no outline of its own —
the only edge it makes is where one color stops and the next begins. Where the atlas
draws no boundary between several names, those names share a color rather
than being split by one, because a color change is a boundary and that split is not a
line this atlas has: on plate 30 the mediodorsal thalamus is one patch, on plate 54 the
cerebellar lobules it prints inside one line are one patch. The faces the atlas seals
and names nothing inside are left unpainted, because there is no region there to
color. Five colors would do 58 of the 62 plates and six the other four; the palette
holds seven, and the spare ones go to the regions that ask for them.
Your own frame
Rotation is experimental. Pitch, roll and yaw are new and not yet fully
tested. Check any coordinate they adjust against anatomy you already know before relying on
it. Moving the origin is not in that category: it is an exact subtraction.
The atlas is cut perpendicular to the brainstem axis, which is not how a head sits in any
particular stereotaxic frame. Frame in the header takes a pitch, a roll and a yaw
about a pivot you choose, and restates the structure card, the pointer readout, the
coordinate lookup and the CSV export in that frame, keeping the atlas figure beside each
one.
It can also be told where zero is, and that part stands on its own. Name bregma,
lambda, the interaural line or the occipital crest and give an offset from it — the
offset is how a zero that is not quite on a landmark gets said, such as half a millimeter
behind lambda — and every coordinate is measured from that point, with the readouts
labeling AP by the landmark it now comes from. Moving zero moves no point, so the
projections read their axes from it too and the measure tool needs no caveat; leave every
angle at 0 and that is all the frame does.
The landmark carries an AP and nothing else, because that is all the atlas prints for
it — but a Height button beside the offset will fill DV in from the fitted
skull. That is what the interaural line needs to be usable as a zero: it is the
ear-bar axis rather than a point on the brain, about 9 mm ventral to the dorsal
plane the atlas measures DV from, so with the height set the readout becomes ordinary
interaural coordinates — AP behind the ear bars, DV up from them. The same row
under the Pivot puts a rotation on that axis instead of on a parallel one
several millimeters too high. The APs stay the atlas’s own and exact; the heights
inherit the skull fit’s approximation, and say so.
A rotation is the part that has to be hedged. While one is set, the grid and the measure
tool stay in atlas coordinates and say so. Note too that a plate then stops being a single
AP: at 17° of
pitch the AP of a point drifts about 2.5 mm between the top and the bottom of the
brain, which is why the readout gives an AP per point. It is a rigid rotation of the
published numbers and not a resectioning — no oblique plate is drawn, or
could honestly be drawn, from a series sampled 350 µm apart.
The projection and the 3-D view are the two that can do more than disclaim a rotation,
because both draw points rather than pixels: an In frame box appears in the
toolbar of each as soon as an angle is set, and ticking it turns the cloud and the stack
into your frame, so up on the screen is your frame’s DV and the axes are the ones
the manipulator drives. One tick serves both — they are two pictures of the same
brain. Untick it and they are back in the atlas’s orientation, which is the one
every published figure is in. Turning a view is still not a resectioning: the same 62
coronal sections are stood up at an angle, not recut. The skull outline and the landmark
rules come along, but neither is simply turned: the outline of a turned skull is not the
turned outline of one, so the fitted skull is flattened again at your frame’s
angle, and a landmark’s coronal plane, once tilted, is drawn where it cuts the
midline (sagittal) or the horizontal through the landmark (top-down) — on the vault,
or for the interaural line the ear-bar axis.
Planning a track
Experimental. New, and not yet checked against a track anybody has
actually driven. Read a plan against anatomy you know before you rely on it.
Plan a track in the side panel takes a structure, a hemisphere and three
approach angles, and gives the entry point, the angles to set on the manipulator and the
distance from the brain surface to the target — drawn live on the plate, the
projections and the 3-D view, and carried in the link so a plan can be sent to someone
or reopened later.
The angles are in your working frame , not the atlas’s. That is the point of
them: the atlas is cut perpendicular to the brainstem axis, and an angle quoted in atlas
coordinates is one nobody can dial into a manipulator. Set a frame above and the plan
moves with it. Because the track is put back onto the plates to be drawn, the picture
stays exact in any frame — this is the one tool here that a rotation does not force
into atlas coordinates.
Tilt drives the tip anterior, roll drives it to the right, and yaw
turns the heading those two aim in — applied in that order. A straight probe spun
about its own axis points where it pointed before, so with tilt and roll both at 0 the
yaw has nothing to turn and the app says so rather than leaving a control that will not
move.
The brain surface is the outline of the section on the nearest plate, traced off
the atlas’s own drawings by the same flood fill the 3-D view runs. It reaches
DV 0 and does not cross it, which is where the atlas defines that plane, and 97.8% of
the printed labels fall inside it. But it is the outline of a fixed, sectioned
brain, not the surface under intact dura, and the sections are 350 µm apart, so an
entry AP is quantized at that spacing however smooth the line looks. Nothing is
interpolated between plates.
And the target is the median position of the printed abbreviation , not the
structure’s centroid — the caveat that runs through everything here, and that
matters most in a tool that invites this much trust. The plan knows nothing about
vasculature, the sinus or the ventricles, and does not check the entry against the skull.
Which plate, and how far off the label. A structure that spans a dozen sections is
labeled on each of them, and the median of all those labels is a point somewhere in the
middle of the structure rather than a point on the section anybody is aiming at.
Take the label from names one plate and reads that section’s labels alone;
picking a plate also turns the viewer to it. Below the angles, the three offset
boxes move the aim off the label — 0.2 mm dorsal to VO on plate 13, say —
and the plan then prints the label and the target as separate rows, so the point the
numbers were measured from stays readable beside the point they are about.
The offset is in atlas millimeters , not the working frame’s, and this is the
one place in the planner where that is true. Naming a target is anatomy: dorsal
has to mean dorsal in the brain rather than up the manipulator’s own axis, or an
offset would mean something different at every pitch. Lateral is taken toward whichever
hemisphere is chosen, so the two sides stay mirror images. Everything downstream —
the entry, the angles, the drive — is still in your frame; the boundary between the
two is exactly the point where the target is named.
Caveats
System tags (auditory, hippocampal, …) are a convenience layer added on top of the
published nomenclature, not part of the original atlas, and the gross divisions are the same kind of addition and a larger one — which structures make up “the pons” is a judgement made here. A structure listed for a plate range is
present at those levels; it is not necessarily labeled on every plate. Label positions come from
OCR — where one was not located, the app says so rather than showing nothing.
Not every name has an outline. The regional outlines are cut from the atlas’s
own drawing, and the drawing sometimes puts several names inside one boundary and
prints nothing between them — the cerebellar lobules and the white matter running
through them, the mediodorsal thalamus, the lateral hypothalamic zones. Those
structures are still found, listed, measured and modeled, but no outline is drawn for
them: hovering or selecting one highlights every place the plate prints its name
instead, and the tip says why. An outline there would be a boundary this atlas does not
have, and a dashed one would still read as a boundary. Where the atlas does
draw the boundary but faintly enough that the extraction had to infer part of it, the
outline is drawn dashed and the tip gives the share of it that is really printed.
The 3-D view
The plates are already in one common coordinate frame, so they can simply be stacked
where they belong: each section is drawn at its own bregma. The view opens on the
Nissl because tissue is what a volume is made of: density reads as density,
where the labeled drawing’s contour ink and printed abbreviations stack into a
haze with a brain somewhere inside it. The plate view still opens on the drawing, and
either can be switched, which is the pairing having two views is for.
Volume , which is what the view opens in, ray-marches that stack and shows
the brain as a solid;
Contours draws the sections themselves, one textured plate at a time, which is
the honest picture of what the stack is — 62 sections and the gaps between them
— and at the default angle 62 cards seen nearly edge-on, which is why it is not
the first thing you see. Labels plots all 6,336 printed abbreviations as a
stereotaxic point cloud — the
projection views with the third axis put back. Nothing extra is downloaded; it is built
from the plate images already on the page, the first time you open it — and once
more for a second staining, since the staining belongs to the pane: Split can
hold a Nissl stack beside a myelin one, the same 62 levels stacked twice and turned
together, which the atlas itself can only offer as two pages you turn between.
Ortho swaps the perspective camera for a parallel one: nothing is
foreshortened, so equal lengths read equally wherever they sit in depth, which is
what you want when comparing positions rather than looking at the brain.
View puts the camera on an anatomical axis — left, right, rostral,
caudal, dorsal or ventral — and a link copied from a named view opens on it.
Caudal is the one that reads the same way round as the printed plates, which are
drawn as seen from behind.
Two things to keep in mind. Sections are 350 µm apart but a plate pixel is about
17.5 µm, so sampling along the brain is 20× coarser than across it: the
stack really is discrete, and the streaking in the volume view is interpolation between
slices, not anatomy. And a label point marks where an abbreviation is printed ,
which for thin layered structures is often just outside the structure itself. Most
structures carry only a handful of points — six is the median — so a single
structure reads as a sparse arc rather than a shape. Nothing here is a segmentation.
Skull draws a CT skull around the atlas: a translucent shell in
3-D, a trace on the current plate, an outline around either projection.
Experimental — the alignment is approximate.
The atlas prints no skull, so the fit is computed rather than published. The scan is a
different animal, aligned first on the paper’s own landmarks (the coronal suture
for bregma, the ear canals for the interaural line); the tilt, height and the three
axis scales are then set by the drawn sections of all 62 plates, which get the last
word. Because the two animals are not the same shape, the scales differ a little by
axis — the skull is stretched about 3% along AP and 4% across, and shortened 4%
in depth, to seat the vault on the cortex. Good to a few tenths of a millimeter.
Context around the sections, not a surface to measure against.
Landmarks marks bregma, lambda and the occipital crest on the plate whose
plane they fall in, and draws the interaural line — solid where it lies in the
plane, dashed elsewhere as the ear-bar height. On the projections all four become
reference rules. In 3-D they stand where they actually sit: each landmark’s coronal
plane as a rule up the midline, a cross on the vault at bregma, lambda and the occipital
crest, and the interaural line as the ear-bar axis it is, running laterally right
through the head and out the other side, with a ring where it passes each canal. That is
the one view where it is a line rather than a point, and it is drawn through the brain
rather than behind it, because an ear bar is not hidden by the head it goes into. The
AP of each is the atlas’s own printed figure and exact ;
only the heights, and the ear canals the rings mark, come off the
skull fit, so they inherit its approximation. Where the bar ends is a drawing choice
and nothing measured: it is carried clear of the widest bone so it reads as a bar going
into a head rather than a chord inside one.
Full method .
The SVG export
SVG saves the same sheet PNG does, but with the plate as vector paths
instead of a photograph: the regional outlines, traced off the printed plate, with
whatever overlays were showing and the same caption. Every part is a named group
— outlines-solid, outlines-dashed,
region-colors, grid,
skull, landmarks, highlight,
measure, track, query, notes,
caption — so a figure can
be restyled or cut down without redrawing anything. About 10–80 kB.
The tracing was done on the published pages at 3296 × 2481 px and
is kept in that frame; each plate carries the matrix that puts it on the
1100 × 703 crop everything else here measures in. That matrix is not
assumed from the page layout — it is fitted to each plate’s own printed ink,
which is what catches the pages the journal set sideways or off-center — plate 20
is printed at a quarter turn, and plate 31 comes out 47 px low in the plate frame,
the same displaced figure the plate crops already correct for. Every plate lands within
1.8 px of its own drawn lines, about 0.03 mm, so the outlines sit where the
grid and the measure tool say they do. The section image is not in the file :
this is the drawing, not the histology.
Getting around
← → step through plates
+ − 0 zoom in, out, and back to fit
drag in 3-D orbit — which leaves any named view; shift-drag or
right-drag to pan, wheel to zoom
Home End the first and the last plate
/ jump to the search box
? open this panel
C the view’s controls — everything the bar above the
picture does not have room for. They dock beside the view, as a column where there
is width for one and a row above it where there is not, so the picture is re-fitted
around them rather than covered. Advanced inside them holds the contrast, the
tone curve and the slab: the settings you make once.
two fingers pinch the plate to zoom, anchored between them; pinch back in
to fit. One finger drags once you are in. The − / + / Fit buttons are
there for a mouse, on a window wide enough to spare them the room.
i on the picturewhat this view is showing, and how to read it
F give the view the whole window — the search column, the
header and the browser’s own chrome go, and the plate, the projection or the
stack takes the room. Esc comes back.
Compare a second plate beside the first, sharing zoom, pan and crosshair
Color regions the whole section filled, no two neighbors alike — the
colors tell them apart and mean nothing else
Notes your own markers, kept in this browser; export them as JSON to keep them
Meshes in 3D, the structure as a closed surface interpolated between its plates
Esc come back from the maximized view, drop the selection, or clear a measurement
double-click zoom into the plate at that point; drag to pan
hover a label read its full name; click it to select that structure
Ctrl -clickadd a region on the plate, or a row in the results, to
the selection, or take it out; ⌘ -click on a Mac. Each is outlined on the
plate, the card stays on the first
This build
Spotted an error in the atlas data, hit a bug, or want a new feature?
Raise an issue on GitHub .
Build {{BUILD_HASH}}{{BUILD_DATE}} . This copy carries no build stamp, so it cannot name the build it came from; the atlas data in it was last recomputed