[Mristudio-users] Relationship between acquisition and reconstructed voxel size.

[Susumu Mori]

On Fri, Aug 16, 2013 at 11:54 PM, Dorian P. <alb.net at gmail.com> wrote:

> Dr. Mori, thank you for the exhaustive explanation.
>
> I have two follow-up questions.
>
> First, you wrote:
> "Because field-of-view, which is determined by the brain size, is about
> 200 - 250 mm, DTI spatial resolution is limited to 2-2.5mm."
> It appears that the number of acquired points depends on the FOV. Does
> this mean that a smaller FOV have more SNR than a bigger FOV for the same
> voxel resolution?
>

Suppose your FOV is 256mm. If the matrix is 96, the voxel size is 2.5mm. If
you make FOV 192mm, the voxel size becomes 2mm. The smaller voxel size
means higher resolution (even though the matrix size is the same) and lower
SNR (less water molecules in each voxel and thus weaker signal). If you
keep the FOV the same (256mm) and increase the matrix to 128, you can also
get 2mm resolution. Again, higher resolution and lower SNR. The former
(192mm/96) is better than the latter (256mm/128) because the data
acquisition time is shorter (acquisition is completed while there remain a
plenty of signal left). However, FOV = 192 mm may be too small for some
subjects. There is a good chance that 192/96 has higher SNR than 256/128.
Also, smaller matrix size allows shorter TE, which contributes to better
SNR. Image distortion is smaller too.


>
> Second, I usually acquire with FOV=230mm. If smaller FOV gets more SNR
> (depending on answer of 1st question), I can reduce FOV to match the
> patients brain. Question is, will variable FOV be problematic in a
> scientific publication?
> I think that a fixed FOV keeps SNR standard between subjects, but brain
> size affects DTI anyway. For example, within the same 2mm voxel size
> chances are that a small brain will have more heterogenous fibers for each
> voxel than a bigger brain. Therefore having more SNR for a smaller brain
> may (kind of) balance the disadvantage. My plain thoughts anyway.
>
>
>
Higher SNR is not due to smaller FOV, but due to smaller matrix size.
192/96 gives noticeably better image quality (SNR, distortion) than
256/128. However, this also means, you should not mix data with different
FOV/Matrix size within one study. I want to use 192/96, but can't because
192mm is too small for some subjects. That's dilemma.

When we do pediatric MRI, there is a big issue; should we keep geometrical
resolution the same or anatomical resolution the same. If we use 2mm
resolution for all subjects, we have the same geometrical resolution. If we
always put 96 voxels from the left edge to the right edge of the brains for
all subjects, we keep the same anatomical resolution. If the subject brain
is 256 mm large, it is 2.5mm resolution and 192mm brain would be 2mm
resolution. We can argue that the latter approach makes more sense from a
biological point of view. However, from a physics point of view, the former
makes more sense because the SNR is kept constant. We know that SNR has a
large impact on FA (lower SNR increases FA).

The conclusion is, there is no perfect study design.


> Thank you for your comments.
>
> Dorian
> TJU
>
>
> 2013/8/16 Susumu Mori <smoriw at gmail.com>
>
>> MRI raw data is so-called time-domain data. This means, within about
>> 10-100ms of time, signals are acquired and recorded. For example, if your
>> image matrix is 128x128, there are 16,384 data points to acquire. For DTI,
>> to freeze the motion effect, all 16,384 points are acquired at once (8,192
>> points if you use a parallel imaging with factor = 2). Actually there are
>> real and imaginary data points and therefore there are 16,384x2 points. By
>> the way, because all 16,384 points are acquired at once, the data
>> acquisition time becomes very long and there is not much signal left by the
>> time all 16,384 points are read from the signal. Therefore, for DTI, there
>> is no point to acquire 256x256 (=65,536) points because after about 20,000
>> point-read, all the remaining points are reading just noise. This is why
>> all DTI studies have been done using 96x96 or 128x128. Because
>> field-of-view, which is determined by the brain size, is about 200 - 250
>> mm, DTI spatial resolution is limited to 2-2.5mm.
>>
>> Now, when we do interpolation by the scanner, the scanner simply add "0"
>> and extend the 128 points to 256 points. This is called zerofilling.
>> After the fourier transform, the time-domain data is converted to the
>> frequency-domain or image-domain data (the same thing with different
>> names).
>>
>> If you have 128x128, after the fourier transformation, you get a
>> 128x128-pixel image.
>>
>> Now, you have two options, do the interpolation by the scanner
>> (time-domain interpolation), convert the 128x128 time-domain matrix to
>> 256x256 time-domain matrix and FT it to the 256x256 image-domain matrix.
>>
>> Alternatively, you can FT first, get a 128x128 image-domain matrix and
>> then digitally interpolate it to 256x256.
>>
>> A big question is, are they different? Signal processing people say,
>> "sinc-interpolation of 128x128 image to 256x256 image is the same as
>> time-domain zerofilling". However, things are not that easy because
>> time-domain data has real and imaginary parts. If you have 10 physicists,
>> I'm sure that you get two camps; one say they are the same and the other
>> say, time-domain interpolation is better and you can never get the same
>> quality after the time-domain data are converted to an image.
>>
>> There is a famous paper by a novel laureate, Dr. Ernst, proving the
>> latter is the case, but the problem is, ordinary people like us can't
>> understand the paper.
>>
>> Anyway, answering your question, you can always interpolate the data into
>> higher resolution, but many people believe that it is just cosmetic,
>> especially in the image-domain. We can argue that you are wasting the hard
>> drive space. I always do twice time-domain zerofilling (128 becomes 256) by
>> the scanner. In the imaging domain, we further digitally interpolate to
>> 1x1x1mm because that is the voxel size of many atlases.
>>
>> I don't think there is a large impact on SNR by the interpolation.
>>
>>
>> On Mon, Aug 12, 2013 at 12:40 PM, Dorian P. <alb.net at gmail.com> wrote:
>>
>>> Hi all,
>>>
>>> Is there any relationship between acquired and reconstructed voxelsize.
>>> Is there any downside of reconstructing to much smaller voxels, for example
>>> acquire at 3mm and reconstruct at 2mm or 1mm? Is SNR going to be the same?
>>>
>>>
>>> Thank you.
>>> Dorian
>>> TJU
>>>
>>> _______________________________________________
>>> mristudio-users mailing list
>>> mristudio-users at mristudio.org
>>> http://lists.mristudio.org/mailman/listinfo/
>>> Unsubscribe, send a blank email to:
>>> mristudio-users-unsubscribe at mristudio.org
>>>
>>>
>>
>> _______________________________________________
>> mristudio-users mailing list
>> mristudio-users at mristudio.org
>> http://lists.mristudio.org/mailman/listinfo/
>> Unsubscribe, send a blank email to:
>> mristudio-users-unsubscribe at mristudio.org
>>
>>
>
> _______________________________________________
> mristudio-users mailing list
> mristudio-users at mristudio.org
> http://lists.mristudio.org/mailman/listinfo/
> Unsubscribe, send a blank email to:
> mristudio-users-unsubscribe at mristudio.org
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.mristudio.org/pipermail/mristudio-users/attachments/20130817/033a8852/attachment.html