[Mristudio-users] DiffeoMap Interpolation Accuracy

[susumu mori]

Hi Jeff,

First of all, if the structure of your interest (in this case SN) is
not well defined (too small or no contrast) in your original image,
computer algorithm cannot define it reliably. Also, the small size
(with respect to pixel size) of the structure would lead to partial
volume effect in raw data (thus poor reproducibility) and, again,
computer algorithm cannot get the reliable information even after
upsampling or normalization. So, everything is in your raw image. I
guess we all agree on it.

To answer your question, maybe a good approach is to use Hmap and warp
the WMPM to your raw image without interpolation. In this way, you may
get better idea about what the automated method is doing more
intuitively. Suppose a ROI is only 10 pixel large in the 1mm WMPM
which is warp to a raw image with 10 times lower resolution. Then, you
will find a ROI with only 1 pixel. This should give you some idea to
judge if the operation is reliable or not. The best way is to scan the
same person multiple times and measure the SN to find the reliability
range.

On Sat, Mar 26, 2011 at 12:26 AM, Jeff Sadino <jsadino.queens at gmail.com> wrote:
> Thank you Hangyi and Susumu for your quick and, as always, detailed replies,
> and my apologies for the unclear question.  Please let me give some more
> details.
>
> Our current workflow is as follows:
> DTI Studio: to create FA, ADC, etc. maps.
> DiffeoMap: AIR registration of FA, ADC, etc., to JHU templates (which
> resamples), then LDDMM registration to JHU templates.  Apply Kimap.
> ROIEditor: Extract FA, ADC, etc on the subject image (now in template space)
> using the Type II WMPM atlas.
>
> One of the ROIs we found significiant was the Nigra.  Our original image
> slice was 4.6mm thick with a .4mm gap.  The Nigra from the Type II WMPM is
> only 6mm thick.  After reading your comments, I would generalize my question
> as: is there a minimum ROI threshold size that I should be using, relative
> to the size of my acquisition voxels?  If we can throw some of the smaller
> ROIs out, it would also help with multiple comparisons.  If it was your
> paper, would you include the Nigra, considering the original slice
> dimensions?  Perhaps this is one area where manual ROI placement would be
> superior to autamated placement.
>
> Thank you very much,
> Jeff Sadino
>
> On Fri, Mar 25, 2011 at 9:20 AM, susumu mori <susumu at mri.jhu.edu> wrote:
>>
>> Hi Jeff,
>>
>> Maybe I didn't quite understand your question, but if you are talking
>> about the LDDMM-based image segmentation, each ROI is 3D. So, I'm not
>> sure if you encounter a situation where you deal with a ROI in one
>> slice, although maybe this is not relevant to your question.
>>
>> Suppose your original image is 3mm and you resample it to 1mm. Then if
>> you draw one 2D ROI, you have a choice of 3 slices within the native
>> 3mm slice. The center slice could be identical to the original image
>> but the edge two slices may be interpolated if you used trilinear
>> interpolation option. If you use the nearest neighbor, you simply get
>> 3 1mm copies of the original 3mm slice.
>>
>> I believe your question can be generalized as "what is the effect of
>> interpolation". Interpolation is a difficult issue. If you have an FA
>> map and interpolate it, you get a different result from interpolating
>> a tensor map and re-calculate a FA map from the interpolated tensor
>> map. There are also multiple options for interpolation. I would
>> recommend you to stick with one method and I don't think that we can
>> conclude one is better than the other.
>>
>> You can also bring the pre-segmented atlas map all the way to your
>> original image space. In this way, you can at least forget about the
>> interpolation issue.
>>
>> Susumu
>>
>> On Fri, Mar 25, 2011 at 1:47 PM, Jeff Sadino <jsadino.queens at gmail.com>
>> wrote:
>> > Hello,
>> > I am wondering about how DiffeoMap upsamples the slice number when
>> > performing linear registration to the JHU template.  After doing the
>> > LDDMM
>> > process, I have some ROIs that look interesting.  However, these ROIs
>> > are
>> > smaller than my slice thickness.  I am wondering how valid it is to draw
>> > conclusions about an ROI that exists entirely at sub-native voxel
>> > resolution.
>> > Thank you,
>> > Jeff Sadino
>> > _______________________________________________
>> > 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
>
>