[Insight-developers] Analyze IO

[Stephen R. Aylward]

Hi Hans,

According to the ITK survey, 45% of ITK users are using ITK outside of 
the medical field.
http://www.itk.org/Wiki/2006_Survey_Summary#Areas_of_Use_of_the_Toolkit

If ITK only did anatomic images, I'd probably agree that an undefined 
image orientation is not needed - but the solution to the lack of an 
orientation should probably be fixed in the readers and image sources 
and not in the writers.

For me, the problem came about because of microscopy work that I began 
at UNC.  They asked for some changes, I did them, recompiled using the 
latest ITK, and now the programs no longer work for them.   They can no 
longer create analyze formatted images which are required by another 
program in their pipeline.

The changes to itkAnalyze mean that we can no longer use analyze for 
microscopy images, for data from any other domain for which image 
orientation is (and should be) undefined, for images that originated as 
stacks of jpegs, etc in which orientation is undefined, for 
phantom/simulated data studies, etc.

Stephen

Hans J. Johnson wrote:
> Stephan,
> 
> The problem with this is that it introduces the very real possiblility of
> getting right/left swaps in your images.  The ITK implementation was written
> to meet the file format definition as specified by the original developers
> (i.e. Mayo clinic as part of the Analyze program).   The bug fixes occurred
> close to the time when the switch from metadata orientation to direction
> cosines was done.  This is mostly because that is when the bug was
> discovered, isolated, and a fix was made.  The original buggy meta-data code
> was written prior to images having a well defined definition concept of
> orientation, and during the implementation of the orientation code (via
> direction cosines) this definition became clear and well defined across all
> file format readers.
> 
> The behavior did change, but it was a bug fix.
> 
> I have very strong feelings on this because analyze files that are written
> to disk with ambiguity have cost me weeks of time tracking down problems.
> Because of this dangerous situation, I believe quite strongly that any file
> format that claims to respect orientation should do so consistently across
> all file formats.
> 
> You proposal below would map the following 3 direction cosign inputs to the
> analyze file filter correctly:
> 
> itk::SpatialOrientationAdapter<3>::DirectionType SAGdir=
> SpatialOrientationAdapter<3>().ToDirectionCosines(itk::SpatialOrientation::I
> TK_COORDINATE_ORIENTATION_PIR);
> SAGITTAL 133124
> 0 0 1
> -1 0 0
> 0 1 0
> 
> itk::SpatialOrientationAdapter<3>::DirectionType CORdir=
> itk::SpatialOrientationAdapter<3>().ToDirectionCosines(itk::SpatialOrientati
> on::ITK_COORDINATE_ORIENTATION_RIP);
> CORONAL 264194
> 1 0 0
> 0 0 -1
> 0 1 0 
> 
> itk::SpatialOrientationAdapter<3>::DirectionType AXIdir=
> SpatialOrientationAdapter<3>().ToDirectionCosines(itk::SpatialOrientation::I
> TK_COORDINATE_ORIENTATION_RPI);
> AXIAL 525314
> 1 0 0
> 0 -1 0
> 0 0 1
> 
> ===============================
> But the other 45 possible orientations of a 3D dataset would also map to the
> RPI definition.  This means that I could only trust SAGITTAL and CORONAL
> image definitions, and would be forced to assume all other definitions are
> "best guesses"
> 
> ###############################
> ###############################
> ###############################
> Another acceptable solution would be to reorient the image data for those
> other 45 possibilities to the RPI orientation prior to writing.  I assume
> that the main problem is that the default direction cosign of
> 1 0 0
> 0 1 0
> 0 0 1
> 
> Is not a legal tag for analyze file formats.
> 
> ###############################
> ###############################
> ###############################
> Finally, there is a bit of a compromise.  I was originally advised not to
> use analyze orient codes of "3,4,5,6" because they are so often mis-used by
> improper 3rd party analyze image implementors:
> 
> typedef enum _analyze75_orient_code
>   {
>     a75_transverse_unflipped = 0,
>     a75_coronal_unflipped = 1,
>     a75_sagittal_unflipped = 2,
>     a75_transverse_flipped = 3,
>     a75_coronal_flipped = 4,
>     a75_sagittal_flipped = 5,
>     a75_orient_unknown = 6
>   } 
>   analyze_75_orient_code;
> 
> Conversations with the Analyze folks indicated that to respect the original
> intent of the file format, that orient codes of "3,4,5,6" were only to be
> used internal to the Analyze program for display purposes, and were never
> supposed to be written to disk.
> 
> I would be quite comfortable with having this implementation put into place.
> I would like to put a prohibition on reading or writing orientaiton codes
> with the unknown orientation flag.  Again force re-orientation of the image
> to a valid (preferably code 0,1,2) orientation before writing.
> 
> Sorry to be so draconian about this, but I have been very embarrassed
> because I depended on ITK getting these orientation done correctly.  When
> using ITK to convert dicom images to both nifti and analyze images, about 5%
> of the 500 images were right left swapped when the nifti and analyze images
> were loaded with ITK tools.  I confirmed the problem by using a simple image
> difference program, and the dicom-nifti always gave 0, but 5% of the
> nifti-analyze showed the right left swap.
> 
> I'm leaving town tomorrow, so won't be able to put this into place, but
> perhaps I can get some help? If this is something that needs to be done
> quickly, Kent Williams should have time during the middle to end of this
> week to work on the changes and tests.  There should be a test to ensure
> that only valid accepted codes can be read and written to disk.  If the
> difference between reading and writing analyze or nifti (or any other
> sufficiently tagged format Meta/Dicom) should be un-ambiguous.
> 
> Ok...enough ranting ;)
> 
> Regards,
> Hans
> 
> 
> 
> On 2/24/07 10:49 AM, "Stephen R. Aylward" <Stephen.Aylward at Kitware.com>
> wrote:
> 
>> Hi,
>>
>> The behavior of the Analyze writer has changed.   Now if the orientation
>> of the image isn't known, an exception is thrown.
>>
>> The changes are related to the switch from metadata orientation usage to
>> oriented image usage.
>>
>> I suggest that if the orientation is unknown/undefined, then we throw a
>> warning, but continue under the assumption of RPI (the default used in
>> other cases).   Only if the orientation is explicit and not a supported
>> type should we throw an exception.
>>
>> I've make the changes in my local version and can forward them to anyone
>> with cvs access.
>>
>> Stephen
> 
> 

-- 
=============================================================
Stephen R. Aylward, Ph.D.
Chief Medical Scientist
Kitware, Inc. - Chapel Hill Office
http://www.kitware.com
Phone: (518)371-3971 x300