BrainWeb: Frequently Asked Questions (with Answers)
See this page for information on
the data format of the BrainWeb downloadable files.
See this page.
The noise in the simulated images has Rayleigh statistics in the
background and Rician statistics in the signal regions. The "percent
noise" number represents the percent ratio of the standard deviation
of the white Gaussian noise versus the signal for a reference tissue.
For custom simulations you can choose the noise reference
tissue. For the pre-computed Simulated Brain Database (SBD), the noise
reference tissues were: White-Matter for all T1-s, and CSF for all
T2-s and PD-s.
For a 20% level, the multiplicative INU field has a range of values of
0.90 ... 1.10 over the brain area. For other INU levels,
the field is linearly scaled accordingly (for example, to a range of
0.80 ... 1.20 for a 40% level).
The INU fields (available for download)
were estimated from real MRI scans, so they are realistic. These fields
are not linear, but are slowly-varying fields of a complex shape.
It is all described on this web page.
It is explained on this page.
Throughout this website you may notice discussion on a fuzzy and a
discrete version of the anatomical phantom. A fuzzy phantom will specify a
mixture of tissue classes for each voxel (eg. a voxel might contain 90%
grey matter and 10% white matter). A fuzzy phantom was used to describe
the tissue within each voxel during the simulation process.
A discrete phantom specifies the tissue type with the largest proportion
in each voxel. A discrete phantom was not used in the simulation process.
However a user might use this phantom when comparing simulated data
against the underlying tissue types (eg. a user might have a
classification algorithm which classifies each voxel into one
specific tissue type).
Which anatomical model to use as "ground truth" for validating an MRI
segmentation method? If your method attempts to estimate the partial
volume (that is, the tissue fractions in each voxel) then the fuzzy
model is probably what you want. Otherwise, the discrete model will
probably do...
Yes, since June 2003 we also have these files available for download.
Indeed. The model was constructed from some real MRI scans using semi-automatic methods (details)
and some MRI acquisition artifacts propagated into the result.
Nevertheless, this adds realism to the simulations, as real-world MRI head images can include unwanted objects too.
Short answer: Not really.
Long answer: Our MRI simulator is quite sophisticated -- it produces the
images starting from tissue MR parameters (ie T1, T2, T2* relaxation
times), as well as incorporating realistic partial-volume, noise, and
intensity non-uniformity (bias field) effects. For more details
see this publication:
-
R.K.-S. Kwan, A.C. Evans, G.B. Pike :
"MRI simulation-based evaluation of image-processing and classification
methods"
IEEE Transactions on Medical Imaging. 18(11):1085-97, Nov 1999.
Nevertheless, you could compute yourself the per-tissue intensity
histograms for a given simulated MR image by using the anatomical
model...
The tissue MR parameters (relaxation times) that we used for the
MRI simulations are listed on this page.
The random seed is used for simulating the noise in the images (and it is ignored if the
noise level is 0%). If you are setting a non-zero seed, you will obtain the same noise in the image everytime you use that particular seed (this is due to how our "random" number generator works). If you set the seed to zero, then the noise will be different everytime you run a simulation.
Non-zero seeds can be useful if you want to obtain several simulated
MRI-s which differ in some parameters but not in the image noise.
This is something that you can do with our simulator but you cannot
with a real-world MR scanner!
The six "protocols" available in the custom simulations interface
({AI,ICBM} x {T1,T2,PD}) are simply examples of real-world
pulse sequences (in other words, parameter templates). Selecting one
of them and then clicking on "Set all parameters" will pre-set all the
custom parameters of the "MR pulse sequence" section to the values
specified by that protocol. You can then adjust these parameters if
you wish to experiment, or if you have your own specific pulse
sequence that you want to simulate.
Note that the selected protocol is not actually used in the
simulation: only the actual pulse sequence parameters (slice
thickness, scan technique, TR, etc) are.
These example scanning protocols are named "AI" and "ICBM" after
two data acquisition projects that took place in our lab.
You may notice that the sample "T1" pulse sequence (for example)
is not the same in the ICBM and AI templates. This is because there is
more than one way to obtain a "T1-weighted MR" image... Don't hesitate
to try your own "T1" sequence, it's cheap and easy!
There is more than one way to produce "T1-weighted" MR images. A
conventional SE sequence (TE/TR = 20/400ms) is one way to do it, and
the spoiled-FLASH sequence from the ICBM template is another way to do
it. The SFLASH sequences are
attractive for real-life scanning due to some practical
advantages, like much shorter scan time (note the low TR), less image
noise, and others.
One benefit of our custom simulations feature is that you can easily
(and cheaply) try out different pulse sequences and compare the
resulting MR images!
No, the current simulator software only produces one 3D image per
simulation. You will have to submit multiple requests in order to get
images acquired at different echo times. Moreover, specifying
multiple echo times is only useful for the DSE_EARLY and DSE_LATE
sequences (for which you need to specify two echoes). For the SE scan
technique only one echo time is used -- if you specify more echo times
then only the first one will be used.
Yes, you can obtain real and imaginary data from our MR simulator in
BrainWeb. They are not part of the pre-generated database (SBD), but
can be obtained through the custom simulations interface.
The software can only produce one "image type" per request, so
you'll have to request two simulations in order to get both real and
imaginary parts. Note: for the real and imaginary data to match, all
the parameters should be identically set except "Image Type". If you
set the noise level to something other than 0%, then you'll have to
also set the "Random generator seed" to some non-zero value (which has
to be the same for the two matching simulations, see also
this question).
If you are downloading the resulting real+imaginary images in a
"raw" format, then there is an extra twist for the two images to be
properly paired: BrainWeb normalizes the output raw data range such
that 0 corresponds to the minimum real value of the simulated image,
and 255 (for raw byte), or 4095 (for raw short), corresponds to the
maximum real value. The min/max real values of the original image are
listed on the download page under "Volume real range". Thus, you need
to appropriately re-scale the real and imaginary parts of such a raw
image pair. (Note: you won't need to worry about this if you are
downloading MINC data, which includes the real range in its header.)
The noise percentage is relative to the average real and imaginary
values of the overall brightest tissue class. Noise is generated using
a pseudorandom Gaussian noise field which is added to both the real
and imaginary components before the final magnitude value of the
simulated image is computed.
Don't hesitate to send us email (bert+bw@bic.mni.mcgill.ca). We will consider adding the requested feature when we have resources available...