Difference between revisions of "NAMIC Wiki:DTI:DICOM for DWI and DTI"
Line 271: | Line 271: | ||
implementing a new mechanism and navigating through a more general | implementing a new mechanism and navigating through a more general | ||
structure. | structure. | ||
− | |||
− | |||
== Related Links == | == Related Links == |
Revision as of 20:55, 25 February 2019
Home < NAMIC Wiki:DTI:DICOM for DWI and DTIThis page should serve as a place where information about DICOM and DWI/DTI data can be maintained. With time, this information could be used as part of automated solutions for learning all the necessary DWI-related information from a DICOM series. A collection of tools for DICOM is here.
As long DICOM support for DWI information is vendor-specific and/or non-conformant with the info here, the Nrrd format provides a means of recording the DWI-specific information once it is known.
Contents
DICOM for DWI
The recommended tags to use in DICOM are as follows:
0018 9075 CS 1 Diffusion Directionality 0018 9076 SQ 1 Diffusion Gradient Direction Sequence 0018 9087 FD 1 Diffusion b-value 0018 9089 FD 3 Diffusion Gradient Orientation 0018 9117 SQ 1 MR Diffusion Sequence 0018 9147 CS 1 Diffusion Anisotropy Type
These are defined in Supplement 49. In particular see section C.8.12.5.9 "MR Diffusion Macro" on pages 94 and 95.
The tags are also referenced in http://medical.nema.org/dicom/2004/04_06PU.PDF (see pages 28-29) as well as in some Working Group Minutes (see pages 155-156).
Two points of interest relative to the NRRD format:
- The definition of "Diffusion Gradient Orientation" implies that the measurement frame is exactly the identity transform.
- There appears to be no means of recording the full B-matrix when it is known. This is not an issue for any NAMIC datasets, but can arise in small-bore imaging.
Private vendor: GE
For GE scanners, Signa Excite 12.0 and later, the following tags are reserved for diffusion weighted images:
- (0019,10e0) : # DTI diffusion directions (release 10.0 & above)
- (0019,10df) : # DTI diffusion directions (release 9.0 & below)
- (0019,10d9) : Concatenated SAT {# DTI Diffusion Dir., release 9.0 & below}
- (0021,105A) : diffusion direction
- (0043,1039) : Slop_int_6... slop_int_9: (in the GEMS_PARM_01 block)
- 6: b_value
- 7: private imaging options 2
- 8: ihtagging
- 9: ihtagspc
This information can be found in http://www.gehealthcare.com/usen/interoperability/dicom/docs/5162373r1.pdf
Unfortunately the Dartmouth DWI data (from a GE Signa scanner) does not conform to this (nor do they use the nominally standard 0x0018 tags), as can be seen by running:
dcdump S4.100 | & grep \(0x0019,0x10
which includes:
(0x0019,0x10d9) DS Concatenated SAT VR=<DS> VL=<0x0008> <0.000000> (0x0019,0x10df) DS User Data VR=<DS> VL=<0x0008> <0.000000> (0x0019,0x10e0) DS User Data VR=<DS> VL=<0x0008> <0.000000>
so all of the tags which are supposed to store # gradient directions store the value 0! In addition, there is:
(0x0021,0x105a) SL Integer Slop VR=<SL> VL=<0x0004> [0x00000000]
so the supposed representation of diffusion-direction is also empty. The Dartmouth data has the following tags describing the scanner and software version:
(0008,1090) LO [GENESIS_SIGNA] (0018,1020) LO [09]
In GE DWI images (software version 12.0)
(0008,1090) LO [SIGNA EXCITE] (0018,1020) LO [12\LX\MR Software release:12.0_M4_0520.a]
diffusion directions are stored under the following tags:
(0019,10bb) DS [0.430617] (0019,10bc) DS [-0.804161] (0019,10bd) DS [-0.420008]
these directions are in the image (rather than scanner bore) frame of reference. The interpretation depends on the Phase Encoding direction (0018,1312). If 0018,1312 is "COL" than 0019,10bb refers to the image columns (i), 0019,10bc refers to the image rows (j) and 0019,10bd refers to the slices (k). If 0018,1312 is "ROW" than 0019,10bb refers to the image rows (j), 0019,10bc refers to the image columns (i) and 0019,10bd refers to the slices (k). There are two caveats: first this assumes that your slices are stacked in spatially ascending order (head to foot for axial acquisitions) regardless of the order that they were saved to disk (which appears to vary depending on how the user plans the scan). Second, you need to adjust the polarity of these directions to match the expectations of the software used for fitting.
Be aware that the b_value stored in tag 0043,1039 may be masked. For example, consider series 16 from this archive. This series has images with b_values of both 750 and 1500, but the DICOM tag stores b_values of 1000001500 and 1000000750, as shown in the DICOM dump from image 24 (below). Therefore, if you use 0043,1039 to determine b_value, you should use the modulus of a large number like 100,000.
(0043,1039) IS [1000001500\8\0\0]
Private vendor: Siemens
Siemens Vida XA10 and XA11 systems will use the recommended public tags (described above) when exporting as enhanced DICOM.
Some Siemens systems, including the Vida XA11 exporting to classic DICOM format will populate the private tags (0019,100c) and (0019,100e) for B-value and B-vector respectively. Here is an example:
(0019,100c) IS [1500] (0019,100e) FD [0.41\-0.40-0.81]
A Siemens DICOM Conformance Statement is available at
http://www.medical.siemens.com/siemens/en_INT/rg_marcom_FBAs/files/brochures/DICOM/mr/dcs_trio.pdf
No diffusion related tags specified.
David Tuch has stated (in email from December 21, 2005):
The diffusion gradient information and coordinate frame are not provided in the DICOM hdr for the MGH diffusion sequences.
Tag 0029,1010 may include all necessary information
If you have installed spm (and Matlab), the following spm-File extracts the gradient info:
P= spm_get(Inf,'*','Selct some files') hdr=spm_dicom_headers(P) hdr{1}.CSAImageHeaderInfo(22).item(1).val hdr{1}.CSAImageHeaderInfo(22).item(2).val hdr{1}.CSAImageHeaderInfo(22).item(3).val
look for "spm_dicom_headers.m" (google or on your hard disk), this spm File shows you how to decode the tag data.
It is easy to write a C or C++ program that does the same, if you have the spm_dicom_headers.m
Credits: Jan Klein <klein AT mevis DOT de>
Update
As far as the latest MR scanner software (2006) version is concerned, a solution for the access of the following Diffusion attributes was provided:
0019;000A;SIEMENS MR HEADER ;NumberOfImagesInMosaic ;1;US;1 0019;000B;SIEMENS MR HEADER ;SliceMeasurementDuration ;1;DS;1 0019;000C;SIEMENS MR HEADER ;B_value ;1;IS;1 0019;000D;SIEMENS MR HEADER ;DiffusionDirectionality ;1;CS;1 0019;000E;SIEMENS MR HEADER ;DiffusionGradientDirection ;1;FD;3 0019;000F;SIEMENS MR HEADER ;GradientMode ;1;SH;1 0019;0027;SIEMENS MR HEADER ;B_matrix ;1;FD;6 0019;0028;SIEMENS MR HEADER ;BandwidthPerPixelPhaseEncode ;1;FD;1
That does not solve your problem with the older datasets and unfortunately there is no easy way to access diffusion information there as it is really only stored in the Siemens shadow part.
Update September 2010
Via email, Matthew Brett reported the following information and links which are posted here with his permission.
...if you do this: P= spm_get(Inf,'*','Selct some files') hdr=spm_dicom_headers(P) hdr{1}.CSAImageHeaderInfo(22).item(1).val hdr{1}.CSAImageHeaderInfo(22).item(2).val hdr{1}.CSAImageHeaderInfo(22).item(3).val from http://www.na-mic.org/Wiki/index.php/NAMIC_Wiki:DTI:DICOM_for_DWI_and_DTI#Private_vendor:_Siemens - then the resulting q-like vectors can be rather far off, for reasons that I now can't remember, but I believe Guy investigated. In SPM, hdr{1}.CSAImageHeaderInfo is the SPM reading of the Siemens private header. We read this in: http://github.com/matthew-brett/nibabel/blob/master/nibabel/nicom/csareader.py in the 'read' function. What SPM calls hdr{1}.CSAImageHeaderInfo(22) has the label 'DiffusionGradientDirection', and is what looks like the q-vector. However, it can be rather far off, and, in particular, is often rather far from a unit vector. Guy recovered from this by using the correctly specified 'B_matrix' field of the same private header. I think this is: hdr{1}.CSAImageHeaderInfo(78) in SPM-speak, and is arranged as (0-based matrix indices into B matrix) (0,0), (0, 1), (0,2), (1, 1), (1, 2), (2, 2) where the B matrix is symmetric. See 'get_b_matrix' in http://github.com/matthew-brett/nibabel/blob/master/nibabel/nicom/csareader.py We (this is obviously optional) rotate to voxel orientation (rather than DICOM orientation). We (this is also optional) (credit Ian Nimmo-Smith) resolve to the nearest positive semi-definite matrix using Niethammer's and C-Fs tensor algorithm - see 'nearest_pos_semi_def' in: http://github.com/matthew-brett/nibabel/blob/master/nibabel/nicom/dwiparams.py Finally, we extract the best estimate of q from the resulting B matrix, using eigenvalue decomposition - see 'B2q' in the dwiparams - this is the main part of Guy's algorithm, and Ian N-S code. But I think you're asking - can I put that on the NAMIC page? I'm hoping that Guy or Ian or Eleftherios will jump in and tell me where I am wrong or explain further...
This bug may or may not be related (now fixed): http://www.na-mic.org/Bug/view.php?id=978
Reference
in particular:
Private vendor: Philips
Be aware that Philips will store derived diffusion images (e.g. isotropic images) in the same series as the raw data. Derived images will disrupt processing. Further, one will be able to generate better derived images after processing (e.g. de-Gibbs, de-noise, eddy correction). Therefore, one typically wants to identify and discard these derived images.
Many Philips images use the following private tags for diffusion weighted images
- (2001,1003) FL : Diffusion B-Factor
- (2001,1004) CS : Diffusion Direction
- (2005,10B0) FL : Diffusion Direction RL
- (2005,10B1) FL : Diffusion Direction AP
- (2005,10B2) FL : Diffusion Direction FH
Note that the possible values for 2001,1004 are P (PreparationDirection), M (MeasurementDirection), S (Selection Direction),O(Oblique Direction), I (Isotropic). Since both the raw B=0 images and derived isotropic use the value "I" this tag is insufficient to discriminate raw images from derived images. For Philips scans, B=0 set all tags 2001,1003; 2005,10B0; 2005,10B1 and 2005,10B2 to zero, while derived images use the corresponding b-value (e.g. 2001,1003 = 1000) but set also set 2001,10B0; 2005,10B1; 2005,10B2 to zero. Raw B-weighted images have a positive value in 2001,1003 and at least one of the gradient directions 2001,10B0; 2005,10B1; 2005,10B2 will be non-zero.
Some Philips images use the recommended public tags described above (for example classic DICOM images from 2012 where 0002,0013 lists Philips MR 26.2 as well as 2018-vintage Philips MR 53.0). Philips enhanced DICOMs also use these public tags.
Complete DICOM conformance statements for
- Intera
- Achieva
- Panorama
- Gyroscan
- Infinion / Eclipse
are available here http://www.medical.philips.com/main/company/connectivity/mri/
Private vendor: Toshiba
The Toshiba Titan 3T (console software V2.30*R005) uses a standard DICOM attribute for b-value:
- (0018,9087) : Diffusion b-value (for instances with a non-zero b-value).
NOTE: this attribute is missing for those instances with a b-value of zero.
Gradients can be found in a private tag as follows:
- (0029, 1001) : Private Sequence
- >(0029, 1090): Private Byte Data
When the byte data is reversed and the nybble-pairs are swapped, this becomes a DICOM sequence object (with 6 items). The 5th item contains a number of standard DICOM attributes including:
- >>(0018,9087) : Diffusion b-value (which matches the above)
- >>(0018,9089) : Diffusion Gradient Orientation. This is the diffusion direction and will be [0, 0, 0] for the b-value of zero instances.
Sample code for the byte data adjustment:
private static byte[] adjustToshibaBytes( byte[] input ) { int len = input.length; byte[] output = new byte[ len ]; for( int i = 0; i < len; i++ ) { output[ i ] = (byte)((input[ len - i - 1 ] & 0x0F) + (input[ len - i - 1 ] & 0xF0)); } return output; }
DICOM for estimated diffusion tensors
There is no specification right now in DICOM to support Tensor. Only a supplement:
Supp 63 Parts 3,4,5,6,16,17 Multi-dimensional Interchange Object
The discussion would then be (D. Clunie quote):
Indeed, even if one were to try to standardize in DICOM the encoding of the entire diffusion tensor, there would no doubt be considerable debate as to whether to do that as 6 (or 9) planes of an "image", since there is such a matrix at each spatial location ("pixel"), or as a special case of the proposed Sup 63 object; the former would keep image-oriented tools and software happier, the latter would require implementing a new mechanism and navigating through a more general structure.