Download Outline Resources Imaging Workflow Medical Image File Formats

Computational Methods in Medical Imaging
Files, DICOM, Geometry
06/04/2011 09:38
Outline
• Patient Imaging.
• File formats.
• DICOM
Reading Image Data Files
DICOM
Geometry & Coordinate Systems
David Atkinson, University College London.
[email protected]
–
–
–
–
–
–
header info
info.
reading images,
anonymisation,
viewing packages,
database,
Enhanced DICOM
format
Resources
• Geometry
– Coordinate systems,
– Radiological image
presentation,
– homogenous coords.
– geometrical
transforms,
– DICOM geometry
information,
– displaying images in
3D.
Imaging Workflow
Web page to accompany these lectures
http://cmic.cs.ucl.ac.uk/staff/david_atkinson/cmimi/
• Image data digitised at scanner
– MRI typically 12 bit ADC, stored as 2 bytes per pixel.
Reference data and presentations
http://dicom.nema.org/
ftp://medical.nema.org/medical/Dicom/Multiframe/
• Transmitted as a series of bytes, saved as a file along with other
information e.g. Patient name.
– Image Pixel data and “header” information.
– Network protocols
Example datasets (Geneva):
http://pubimage.hcuge.ch:8080/
Geometrical Transforms and homogeneous coordinates.
Foley, van Dam, Feiner, Hughes. Computer Graphics. Chapter 5
• Displayed as 2D images or 3D representation.
– Hanging protocols.
– Geometry information
• Radiologist reports, surgeon plans operation, radiotherapy dose
plan calculated.
– Patient records
Medical Image File Formats
Single file
Two files
• Information and image in
one file.
• DICOM, gipl, nifti, ...
• Information in one file,
image in another.
• Mayo Analyze .img & .hdr
• Nifti
• Need to know the file
format....
• Image easy to read.
• Potential to loose one file.
DICOM is a vast set of standards
Focus here on format for medical
image files, especially geometry
information.
[NEMA documentation]
David Atkinson, UCL
1
Computational Methods in Medical Imaging
Files, DICOM, Geometry
06/04/2011 09:38
Non DICOM formats
Reading an Unknown File Format
• Find the size N of the file in bytes.
• Assume the size is dominated by the data from a square
y x y image and this data is at the end.
• The approx. number of pixels is N/b, where b is the bytes
per pixel (typically b=2).
b=2)
• Take the square root and round down to y, a power of 2,
e.g. y=256
• Image is the last y.y.b bytes.
1
• Gipl
– Fixed 256 byte header
– Originates from Guy’s Hospital
– Poorly documented
• NifTi
– http://nifti.nimh.nih.gov/
– Either dual files .hdr & .img, or single, .nii
N
Header
Image Data
y.y.b
Simplified DICOM File Structure
• Many other formats
• Good web site is David Clunie’s and his
Medical Image Format FAQ.
– http://www.dclunie.com/
htt //
d l i
/
information
pixel values
P ti tN
PatientName
ImageOrientationPatient
ImagePositionPatient
PixelSpacing
• Medical imaging dominated by DICOM
…
PrivateFields
DICOM Tags
• ACR-NEMA origin
• Public and Private Fields
DICOM in MATLAB
• dicominfo returns the information in a
structure
• dicomread returns the image data
PatientName
ImageOrientationPatient
ImagePositionPatient
info = dicominfo(‘dicomfile’) ;
X = dicomread(info) ;
PixelSpacing
…
X2 = dicomread(‘dicomfile’) ;
PrivateFields
David Atkinson, UCL
2
Computational Methods in Medical Imaging
Files, DICOM, Geometry
MATLAB example
06/04/2011 09:38
imshow(dicomread('CT-MONO2-16-ankle.dcm'),[])
IEEE 754 float
binary representation
Binary Representation of an Integer
− 1sign × 2exponent −bias × 1 ⋅ mantissa
23 22 21 20
8 4 2 1
1
0
1
1
binary (4 bits)
sign
exponent bias
mantissa
single
1 bit
8 bits
+127
23 bits
double
1 bit
11 bits
+1023
52 bits
8 + 0 + 2 + 1 = 11 (decimal)
uint16 16 bits = 2 bytes, 216 = 65,536
represents 0 = 65,536
212 = 4096
Calculating original pixel values
• Pixels often stored as uint16.
• Use rescale slope and intercept to convert
to original values.
• !!!!! Convert to double first or you may
loose negative numbers e.g. Hounsfield
units, MR phase values, inversion
recovery !!!!!
info = dicominfo(‘myfile.dcm’) ;
X = dicomread(info) ;
S = info.RescaleSlope * double(X)
David Atkinson, UCL
32
bits
64
bits
Anonymising DICOM
How do you effectively anonymise?
Is it the “right”
right thing to do?
– good for data protection / patient
confidentiality
– leads to data and results that cannot easily
be linked back to clinical outcomes etc.
– Identification / recognition still possible.
+ info.RescaleIntercept ;
3
Computational Methods in Medical Imaging
Files, DICOM, Geometry
06/04/2011 09:38
Personal data in DICOM header
• Supplement 55: Attribute Level Confidentiality
(including De-identification) outlines tags and
issues
ftp://medical.nema.org/medical/dicom/final/sup55_ft.pdf
Copying only required info.
• See doc for dicomanon in R2010a or
higher.
• Use my dinfostrip
• Name and birthdate are obvious,
obvious but there can
be others hidden that are not even in this list, or
protected by the MATLAB dicomanon function
e.g. patient address.
• Safest approach is to copy only the image and
attributes that you specify
– resulting file may no longer be fully DICOM compliant.
Managing DICOM Files
• Problem: Thousands of files with
meaningless names. How do you store,
browse and retrieve the ones you want?
• Tools
– OsiriX: Mac based viewer.
– K-PACS: Viewer working under Windows
– DCM4CHE: Database and web viewer.
– Many others – see David Clunie’s pages
David Atkinson, UCL
OsiriX
•
•
•
•
•
•
Mac based only.
Free.
Good functionality.
Popular with clinicians.
FDA approved versions for sale.
Allows user to add and write plugins.
4
Computational Methods in Medical Imaging
Files, DICOM, Geometry
06/04/2011 09:38
K-PACS
•
•
•
•
Easy to download and install.
http://www.k-pacs.net
Free.
Can read a CD or connect to a database
e.g. DCM4CHEE
K-PACS screenshot
DCM4CHE
• Used ~clinically. Many features.
• Supports a web browser.
• Can receive from equipment such as MR
scanners and interfaces with viewers such as
OsiriX and K-PACS.
• Free.
• Tricky to set up with limited documentation.
• Web browser only shows single images.
K-PACS screenshot
DCM4CHE Installation Notes
• Took me about 1 day from scratch to put it on a laptop.
• Follow these instructions:
http://www.dcm4che.org/confluence/display/ee2/Installation
• You will also need (free) :
–
–
–
–
a database server e.g. mySQL
Java Development Kit
JBoss
Patience and Google to get around problems
• Comments on mySQL installation
–
–
–
–
Use the default settings in installation
Allow mySQL in f-SECURE and repeat installation.
Needed(?) to allow port 3360 in f-secure firewalls settings.
Get error in mySQL installation (installer doesn’t set password
apparently) – for fix see http://www.rad89.com/mysql_install.html
Dcm4chee web front-end
David Atkinson, UCL
5
Computational Methods in Medical Imaging
Files, DICOM, Geometry
DICOM old vs enhanced
06/04/2011 09:38
Enhanced attributes for MRI
• Old style
– one file per slice – huge numbers of files.
– Important parameters e.g. diffusion weighting, hidden
in non-standard Private Fields.
• Enhanced DICOM
– multi-frame,
– better information about 3D and time,
– many more parameters in Public Fields (was 2 for
MRI, now 94)
– raw data archive possible.
• Separate gradient
and RF echo train
lengths
• Out-of-plane phase
encoding steps
• Flow compensation
• Spectrally selective
excitation &
suppression
• Blood signal nulling
• Tagging
• Diffusion values and
direction
• Spatial saturation
slabs
l b
• Velocity encoding
• Chemical shift
imaging (metabolite
maps)
[medical.nema.org]
Enhanced DICOM
Organizational Features
• Multi-frame pixel data
• Shared and per-frame functional groups
– Each functional group contains attributes that likely vary as
a group, e.g. Pixel Measures, Plane Orientation, Velocity
Encoding, etc.
– Compact & makes explicit what doesn’t change
• Dimensions
– a priori hints as to how the frames are organized
– Specify intended order of traversal, such as space, then
time (e.g., for cardiac cine loops)
• Stacks
– Groups of spatially-related slices, repeatable
• Temporal positions
[SPIE 2003]
[medical.nema.org]
Organization of Data
• Goal is to reduce the work that the receiving
application has to do to “figure out”
– How the data is organized
– Why it is organized that way
• Without preventing use of the data in
unanticipated ways
– E.g. 3D on a dataset not intended as a volume
• Two levels
– The detailed shared & per-frame attributes
– The overall dimensions, stacks and temporal
positions
[SPIE 2003]
David Atkinson, UCL
[medical.nema.org]
6
Computational Methods in Medical Imaging
Files, DICOM, Geometry
06/04/2011 09:38
Trigger Temporal
Delay Position
Time
Index
Dimensions
Stack ID = 1
Start with a dimension of
space.
A set of contiguous slices
through the heart.
Add dimension of time
(delay time from R-wave).
48 ms
5
2
Sets of contiguous slices
throughout cardiac cycle.
4
3
2
1
In-Stack Position
Stack ID = 1
Stack ID = 1
Time
0 ms
5
1
5
4
4
3
3
2
2
1
[medical.nema.org]
Space
In-Stack Position
[medical.nema.org]
Trigger Temporal
Delay Position
Time
Index
Trigger Temporal
Delay Position
Time
Index
Stack ID = 1
1\5\2
48 ms
1
Space
In-Stack Position
Dimension
Index
Values
5
2
Stack ID = 1
1\5\2
48 ms
4
5
2
4
3
3
2
2
1
1
In-Stack Position
Stack ID = 1
1\5\2
1\4\2
1\3\2
1\2\2
1\1\2
In-Stack Position
Dimension Index Pointers:
1. Stack ID
2. In-Stack Position
3. Temporal Position Index
Dimension Index Pointers:
1. Stack ID
2. In-Stack Position
3. Temporal Position Index
Stack ID = 1
Time (2)
0 ms
1
Time (2)
0 ms
5
1
5
4
4
3
3
2
2
1
[medical.nema.org]
1\5\1
1\4\1
1\3\1
1\2\1
1\1\1
Trigger Temporal
Delay Position
Time
Index
Stack ID = 1
2\1\5
5
2
4
3
2
1
Stack ID = 1
Stack ID = 1
Dimension
Index
Values
2\1\5
2\1\4
2\1\3
2\1\2
2\1\1
In-Stack Position
2\1\5
48 ms
5
2
4
3
2
1
In-Stack Position
Dimension Index Pointers:
1. Temporal Position Index
2. Stack ID
3. In-Stack Position
1
5
4
3
2
1
In-Stack Position
[medical.nema.org]
David Atkinson, UCL
1\1\5
1\1\4
1\1\3
1\1\2
1\1\1
Dimension
Index
Values
2\1\5
2\1\4
2\1\3
2\1\2
2\1\1
Stack ID = 1
Time (1)
0 ms
Space (1)
In-Stack Position
[medical.nema.org]
Trigger Temporal
Delay Position
Time
Index
48 ms
1
Space (1)
In-Stack Position
Dimension
Index
Values
Dimension Index Pointers:
1. Trigger Delay Time
2. Stack ID
3. In-Stack Position
Time (1)
0 ms
1
5
4
3
2
1
Space (2)
In-Stack Position
1\1\5
1\1\4
1\1\3
1\1\2
1\1\1
Space (2)
[medical.nema.org]
7
Computational Methods in Medical Imaging
Files, DICOM, Geometry
06/04/2011 09:38
Dimension features
• Description of dimensions separate from their indices
– Dimensions are described once
– Indices within dimensions are encoded per-frame
• May be multiple sets of dimensions in one object
– E.g., Set
S 1
1: space then time, S
Set 2
2: time then space
• Receiving application only needs to follow the index
values
– Does NOT need to select or sort by attribute value
– Dimensions can be entire functional groups
– Dimensions can be private attributes or functional groups
[medical.nema.org]
Dimension applications
• Selection of sort order for simple viewing
• Partitioning of frames for “hanging”
• Selection of frames that constitute a
– volume in space
– temporal sequence
– contrast administration phase
– physiological parameter, e.g. diffusion b value
[medical.nema.org]
Diffusion
Appears to be structure for recording either the
diffusion gradient direction, or the full b-matrix in the
patient coordinate system.
Multi-frame Diffusion: MATLAB
• Example Philips file
• 13Mb in size, 374 frames
• (13010976 bytes, 128x128 images)
dinfo.NumberOfFrames
374
X = dicomread(dinfo);
size(X)
128
128
1
374
montage(X,’DisplayRange’,[0 200])
Using the dimension feature?
>> dinfo.DimensionIndexSequence.Item_1
ans =
DimensionOrganizationUID:
'1.3.46.670589.11.0.5'
DimensionIndexPointer: [32 36950]
FunctionalGroupPointer: [32 37137]
David Atkinson, UCL
8
Computational Methods in Medical Imaging
Files, DICOM, Geometry
06/04/2011 09:38
Using the perFrame...
Dinfo.PerFrameFunctionalGroupsSequence
ans =
Item_1: [1x1 struct]
Item_2: [1x1 struct]
Item_3: [1x1 struct]
Item_4: [1x1 struct]
Item_5: [1x1 struct]
Item_6: [1x1 struct]
...
Item_374: [1x1 struct]
Complicated info structure
>>dinfo.PerFrameFunctionalGroupsSequence.Item_234.M
RDiffusionSequence.Item_1.DiffusionGradientDirectionS
equence.Item_1.DiffusionGradientOrientation
ans =
0.2626
-0.9548
-0.1389
Unclear how to resolve this issue.
Geometrical Information in
DICOM
Siemens Mosaic
• Unofficial (?) format.
• Each file is a mosaic (montage) of slices.
• One mosaic file per diffusion direction, for
example.
l
• Relatively easy to handle.
• Relatively few files.
Geometrical Transforms in
Therapy
•
•
•
•
•
•
Introduction
Coordinate systems
Homogeneous coordinates
Geometrical transforms
DICOM file information
Display.
Overall registration transform is
composed of a series of stages
Calibration + tracking during
intervention
M1
Scanner Space
3D Image Space
X-ray Table Space
λ
λT
M2
[VisionRT]
David Atkinson, UCL
Dicom image header M1
Calibration M2, T, P
Tracking M3
Patient imaging λ, R
R*P
X-ray C-arm
Space
M3
2D Image Space
[Kawal Rhode, KCL]
9
Computational Methods in Medical Imaging
Files, DICOM, Geometry
06/04/2011 09:38
Right Handed Coordinate
System
and display
x
y
thumb
1st finger
z
Example Scanner and Patient
Coordinate Systems
2nd finger
Radiologist vs Neurosurgeon
!!Left-Right swap in image!!
L Left
P Posterior
H Head
Z/H
X/L
For a supine
(face-up) patient
Neurosurgeon
Head = superior (S)
Foot = inferior (I)
Y/P
[www.itk.org]
Radiologist
Radiographic Presentations
Head
Coronal
Right
Head
Sagittal
Left
Anterior
Posterior
Foot
Anterior
Right
MRI: phase, read, slice
Foot
Gradient directions
can be reversed e.g.
to change the direction
of a fat shift.
This means the prs
coordinate system
could be left or right
handed!
phase
slice
Left
Readout (frequency encoding)
Axial / Transverse
Posterior
David Atkinson, UCL
View patient from Left foot
10
Computational Methods in Medical Imaging
Files, DICOM, Geometry
DICOM helps in the face of confusion
06/04/2011 09:38
How do we place DICOM file in 3D?
• DICOM LPH and Cartesian axes are both
right handed.
• Vectors, matrices, trigonometry and
coordinates all work “as
as expected”
expected in any
right handed system.
• DICOM standard enforces standard
geometry information in the file.
Machinery:
•Coordinate Transforms
•Affine transforms
•Homogeneous Coordinates
•Matrix vector products
•Cross product
Matrices for Coordinate Transforms
• Rotation matrix
transforms
coordinates of a
point by a rotation.
b
yb
ya
a
θ
xb = xa cosθ − ya sin θ
xb xa
yb = xa sin θ + ya cosθ
⎡ xb ⎤ ⎡cosθ
⎢ y ⎥ = ⎢ sin θ
⎣ b⎦ ⎣
− sin θ ⎤ ⎡ xa ⎤
cosθ ⎥⎦ ⎣⎢ ya ⎥⎦
• In 2D, rotations and scales can be
expressed by 2x2 matrices.
• However, translations require a “trick”:
Homogeneous Coordinates.
Coordinates
– Add another dimension with value 1 to the
coordinates.
– Add another row and column to the transform
matrix.
Homogeneous Coordinates
• A convenient way of representing coordinates
that allows for geometric transformations
including translations.
• Add an extra dimension.
w
P
⎡ xb ⎤ ⎡cos θ
⎢ y ⎥ = ⎢ sin θ
⎣ b⎦ ⎣
− sin θ ⎤ ⎡ xa ⎤
cos θ ⎥⎦ ⎢⎣ ya ⎥⎦
⎡ xb ⎤ ⎡cos θ
⎢ y ⎥ = ⎢ sin θ
⎢ b⎥ ⎢
⎢⎣ 1 ⎥⎦ ⎢⎣ 0
− sin θ
cos θ
0
0 ⎤ ⎡ xa ⎤
0⎥⎥ ⎢⎢ ya ⎥⎥
1⎥⎦ ⎢⎣ 1 ⎥⎦
Rotation matrix
Rotation applied to
homogeneous coordinate
xb = xa + t x
y b = ya + t y
1
x
[Foley]
y
David Atkinson, UCL
Translations represented this way.
⎡ xb ⎤ ⎡1 0 t x ⎤ ⎡ xa ⎤
⎢ y ⎥ = ⎢0 1 t ⎥ ⎢ y ⎥
y ⎥⎢ a ⎥
⎢ b⎥ ⎢
⎢⎣ 1 ⎥⎦ ⎢⎣0 0 1 ⎥⎦ ⎢⎣ 1 ⎥⎦
11
Computational Methods in Medical Imaging
Files, DICOM, Geometry
06/04/2011 09:38
Alternative representation:
(row vector) * matrix
Comments
• Homogeneous coordinates allow
representation of affine transformations.
• Transforms can be concatenated into a
series of matrix multiplications
p
((in the
correct order).
⎡ cosθ
ya ] ⎢
⎣− sin θ
[xb
y b ] = [ xa
[xb
y b 1] = [xa
• Take care assembling the matrix:
(row vector)*matrix, or, matrix*(column vector)
Affine Transforms
• An affine geometrical transform keeps
parallel lines parallel.
• Includes:
– translations,
t
l ti
– rotations,
– scale changes,
– skews.
0
⎡1
⎢0 cosθ
Rx = ⎢
⎢0 sin θ
⎢0
0
⎣
⎡cosθ
⎢ sin θ
Rz = ⎢
⎢ 0
⎢ 0
⎣
y b 1] = [xa
0
− sin θ
cosθ
0
0⎤
0⎥
⎥
0⎥
1⎥⎦
0 sin θ
1
0
0 cosθ
0
0
0⎤
0⎥
⎥
0⎥
⎥
1⎦
− sin θ
cosθ
0
0
0
0
1
0
0⎤
0⎥
⎥
0⎥
1⎥⎦
Note matrix transpose!
⎡ cosθ
ya 1] ⎢ − sin θ
⎢
⎣⎢ 0
cosθ
⎡⋅
⎢
ya 1] ⎢ ⋅
⎢⎣t x
0⎤
⎥
0⎥
1 ⎥⎦
⋅
⋅
ty
sin θ
0
0⎤
0⎥
⎥
1⎥⎦
Convention used by
MATLAB maketform
but makehgtform uses
the matrix vector
convention!
3D Affine Transformation Matrices
matrix * (column vector) format
⎡sx
⎢0
scale ⎢
⎢0
⎢0
⎣
0
sy
0
0
0
0
sz
0
0⎤
0⎥
⎥
0⎥
1⎥⎦
⎡1
⎢λ
2D skew or shear ⎢ y
⎢0
⎢0
⎣
… 3D rotations
⎡ cosθ
⎢ 0
Ry = ⎢
⎢− sin θ
⎢ 0
⎣
[xb
sin θ ⎤
cosθ ⎥⎦
λx
1
0
0
0
0
1
0
0⎤
0⎥
⎥
0⎥
1 ⎥⎦
⎡1
⎢0
translation ⎢
⎢0
⎢
⎣0
0 0 tx ⎤
1 0 t y ⎥⎥
0 1 tz ⎥
⎥
0 0 1⎦
Applying transforms to coordinates
• transform applied to
multiple coordinates to
transform an object.
• in homogeneous
coordinates, matrices
can be multiplied to
compose transforms
Positive rotation in a RH coordinate system: when looking towards origin from
positive axis, a 90 degree counterclockwise rotation transforms one positive axis
into the other. e.g. rotating around x-axis, positive rot is y to z.
David Atkinson, UCL
12
Computational Methods in Medical Imaging
Files, DICOM, Geometry
06/04/2011 09:38
MATLAB code
• Compose transforms into one matrix.
• Arrange multiple coordinates into one
Transform
matrix
⎡ xb ⎤ ⎡ ⋅ ⋅ ⋅ ⎤ ⎡ xa ⎤
⎢ y ⎥ = ⎢ ⋅ ⋅ ⋅⎥ ⎢ y ⎥
⎢ b⎥ ⎢
⎥⎢ a ⎥
⎣⎢ 1 ⎥⎦ ⎢⎣0 0 1⎥⎦ ⎢⎣ 1 ⎦⎥
⎡ xb
⎢y
⎢ b
⎣⎢ 1
ub L
vb L
1
1
• coord_transform.m
One coordinate
⎤ ⎡ ⋅ ⋅ ⋅ ⎤ ⎡ xa
⎥ = ⎢ ⋅ ⋅ ⋅⎥ ⎢ y
⎥ ⎢
⎥⎢ a
⎦⎥ ⎣⎢0 0 1⎦⎥ ⎣⎢ 1
ua L
va L
1
1
⎤
⎥
⎥
⎦⎥
Multiple coordinates
Geometry Information in DICOM
DICOM definitions
ImagePositionPatient
PixelSpacing(1)
• DICOM uses a right handed LPH coordinate
system.
• Relates to patient, not scanner, but relies on
radiographer
g p
correctly
y entering
g supine/prone.
p
p
• Origin is arbitrary (not isocentre) but fixed.
• Nifti uses RAH (also right handed)
• Analyze uses LAH (left handed!!)
• DICOM provides public fields that relate a 2D
image to 3D patient space.
Quiz.
ImageOrientationPatient(1:3)
PixelSpacing(2)
ImageOrientationPatient(4:6)
PixelSpacing and ImagePositionPatient are in mm
ImageOrientationPatient are two unit vectors (direction cosines)
Height and Width give number of rows and columns.
Finding an Image Pixel Coordinate
in LPH
IOP(1:3)
IOP(4:6)
• What are the ImageOrientationPatient
vectors for these images (radiological
presentation)?
P
[0 1 0 0 0 -1]
[1 0 0 0 0 -1]
[1 0 0 0 1 0]
H
L
IPP +
3 * IOP(1:3) * PixelSpacing(2) +
1 * IOP(4:6) * PixelSpacing(1)
IPP = ImagePositionPatient
IOP = ImageOrientationPatient
David Atkinson, UCL
13
Computational Methods in Medical Imaging
Files, DICOM, Geometry
06/04/2011 09:38
MATLAB Default Image
Coordinates
Placing a 2D DICOM image in 3D
•
[0.5 0.5]
To use MATLAB surf, we need the 3D
coordinates of the pixel vertices.
• Two possible approaches:
1. Starting with ImagePositionPatient,
use vector addition
2. Define image coordinates in MATLAB
default system, then apply transformation
matrices to get 3D patient system.
[0.5 1.5]
[1 1]
Row
numbers
increase
going DOWN
[4 1]
Image coordinates in [row column] order
Pixel centres are integer coordinates.
Displaying a 2D image in 3D using
surf(X,Y,Z,img)
• 2D matrices X,Y
and Z contain
patient coordinates
of the vertices of
the patches.
• 2D matrix img
contains patch
“colours”.
• Sizes of X,Y and Z
are one greater
than img in each
dimension
X(1,1)
X(1,2)
Y(1,1)
img(1,1)
g( )
img(1,2)
img(2,1)
img(2,2)
Y(3,1)
recap
1) Vector addition
1. Find the 3D patient coord of the top left
vertex.
2. Using vector additions, calculate the 3D
coordinates of all other vertices
vertices.
P
H
L
IPP +
3 * IOP(1:3) * PixelSpacing(2) +
1 * IOP(4:6) * PixelSpacing(1)
IPP = ImagePositionPatient
IOP = ImageOrientationPatient
David Atkinson, UCL
14
Computational Methods in Medical Imaging
Files, DICOM, Geometry
ipp − 0.5 × (arˆ + bcˆ )
Top left vertex in patient coord
a = ps(2)
b = ps(1)
06/04/2011 09:38
2) Matrix transform from Image to
Patient Coords
Building transformation matrices:
• ImageOrientationPatient
– rotation
ipp
• ImagePositionPatient
rˆ = iop(1 : 3)
– translation
• PixelSpacing
– scaling
cˆ = iop(4 : 6)
• Compose matrices and apply to image
coordinates
iop = ImageOrientationPatient
ipp = ImagePositionPatient
ps = PixelSpacing
Direction Cosines
Direction Cosines
x̂
a
L
θx
θy
rˆ = [a b]
rˆ = axˆ + byˆ
r̂
b
r̂
rˆ = a 2 + b 2 = 1
ŷ
rˆ = cos(θ x )xˆ + cos(θ y )yˆ
Rotate
Re-label
The components of a unit vector (a & b here) are the
cosines of the angles the vector makes with the basis
directions
Direction Cosines and DICOM
L
r̂
0
rˆ = [a b]
cˆ = [−b a]
P
Image to Patient Transform:
direction cosines as matrix columns
rˆ = [a b]
cˆ = [−b a]
r̂
(0,0)
⎡a ⎤
⎡a ⋅⎤ ⎡1⎤
rˆ : ⎢ ⎥ = ⎢
⎥⎢ ⎥
⎣b ⎦ LP ⎣b ⋅⎦ ⎣0⎦ xy
P
David Atkinson, UCL
x
ĉ
⎡− b⎤
⎡⋅ − b ⎤ ⎡0⎤
cˆ : ⎢ ⎥ = ⎢
⎥⎢ ⎥
⎣ a ⎦ LP ⎣⋅ a ⎦ ⎣1⎦ xy
y
ĉ
(1,0)
⎡a − b⎤
⎢
⎥ = [rˆ : cˆ :]
⎣b a ⎦
Simplified Image
(here pixel sizes 1,
origin at top left)
15
Computational Methods in Medical Imaging
Files, DICOM, Geometry
Composing the overall transform
from Image to Patient
Direction Cosines in 3D with
homogeneous coordinates
Rotation matrix composed
from row, col. and slice
direction cosines as columns
sˆ = rˆ × cˆ
06/04/2011 09:38
Image
coordinate
M = TIPP R S T0
Shift image to make
top left voxel centre at
(0,0,0)
Patient system
⎡ xp ⎤
coordinate
⎡ rdcx cdcx sdcx 0⎤ ⎡ xim ⎤
⎢ y ⎥ ⎢rdcy cdcy sdcy 0⎥ ⎢ y ⎥
p
⎢ ⎥=⎢
⎥ ⎢ im ⎥
⎢ z p ⎥ ⎢ rdcz cdcz sdcz 0⎥ ⎢ zim ⎥
⎢ ⎥ ⎢
⎥⎢ ⎥
0
0
1⎦ ⎣ 1 ⎦
⎣1⎦ ⎣ 0
Scale using PixelSpacing
Rotate into Patient coordinate
system using Direction Cosines
from ImageOrientationPatient
Translate to put top left pixel centre at ImagePositonPatient
[See http://www.electromagnetics.biz/DirectionCosines.htm]
Applying the transform to multiple
coordinates “at once”
⎡ l1
⎢p
⎢ 1
⎢ h1
⎢1
⎣
l2
p2
h2
1
L⎤
⎡ x1
⎥
⎢y
L
⎥ = M⎢ 1
L⎥
⎢ z1
⎥
⎢
⎦
⎣1
x2 L⎤
y2 L⎥
⎥
z 2 L⎥
⎥
1
⎦
Stacking Slices
Problem: Multiple 2D slices, each stored as
a separate DICOM file – how do you
assemble into a 3D matrix?
y on file naming.
g
• Do not rely
• Find the through-slice direction using the
vector product n = IOP(1:3) x IOP(4:6)
• For each file, compute the component of
IPP in this through-slice direction (n.IPP)
and sort.
David Atkinson, UCL
Having calculated the coordinates of the vertices, use surf
to display.
Allows interactive spinning of 3D plots.
plots (Camera toolbar)
Cross Product gives slice
normal
• Cross product of IOP vectors gives
through slice direction, correctly in a RH
coordinate system.
+ve ?
16
Computational Methods in Medical Imaging
Files, DICOM, Geometry
Geometry Fields in Multi-frame
DICOM
ImageOrientationPatient
SharedFunctionalGroupsSequence.Item_1.PlaneOrientationSequence.Item_1.ImageOrientationPatient
I
ImagePositionPatient
P iti P ti t
PerFrameFunctionalGroupsSequence.Item_168.PlanePositionSequence.Item_1.ImagePositionPatient
(ignore specific Item numbers here)
Comments
• ImageOrientationPatient,
ImagePositionPatient and PixelSpacing
are enough to specify a slice in 3D.
• Take care putting slices into a 3D file format
such as gipl, Analyze, Nifti. Compute the
through-slice direction from the cross product of
the row and column directions. Sort slices along
this direction.
• Ignore file alphabetical ordering e.g. from dir
06/04/2011 09:38
Diffusion Gradient in multi-frame
dinfo = dicominfo(filename);
dinfo.PerFrameFunctionalGroupsSe
quence.Item_10.MRDiffusionSeque
10
iff i S
nce.Item_XX.DiffusionGradientDi
rectionSequence.Item_1.Diffusio
nGradientOrientation
Coordinate and Rotation Origins
• Coordinate origin usually does not
coincide with image centre.
• Affine R matrix assumes rotation about
g
coordinate origin.
• Recipe: Translate image such that origin is
at centre, apply rotation, translate back.
These can be composed, note order is
important.
Fibre orientation in diffusion
• Easiest is to stick in the patient system
and use vector addition to IPP.
• In this scheme, you do not need to worry
about MATLAB vs image coordinates
coordinates,
vertices, rotations etc
David Atkinson, UCL
• Compute tensor from Diffusion weighted images and
gradient directions.
• Compute principal eigenvectors.
• Display or track these vectors.
17
Computational Methods in Medical Imaging
Files, DICOM, Geometry
06/04/2011 09:38
Fibre orientation in diffusion
• The eigenvectors are in the gradient coordinate system.
– In DICOM coords for newer files.
– Otherwise, could be physical gradients or phase, read, slice
which rotate with the slice orientation..
• The image is probably displayed in the DICOM
coordinate system.
• The next pixel in the direction of the principal eigenvector
has to be taken from the image volume (stored in the
image coordinate system).
• Check by following fibres around known corners.
Case Study
• Clinician gives you a CD of files, mostly
DICOM but also some image files, notes
etc.
• DICOM files have meaningless names,
names
may not be in slice order and span a
number of scans.
• How do you process these?
Example code
1. Take an initial look using DICOM viewing
software to note scan numbers.
2. In MATLAB, create a structure array with each
element corresponding
p
g to a DICOM file.
3. Use find on the array to locate the relevant
entries for the scan you require.
4. Read in the corresponding images.
5. Using the geometry information, sort slices into
the correct order.
David Atkinson, UCL
• dicomprocess.m processes a DICOM
dir
• isdicom_da.m modification to
MATLAB’s
MATLAB
s isdicom.m
isdicom m which can fail for
very small files.
• dicom2vol.m put slices in order
• Note, this is working code, not optimised
for teaching.
18
Computational Methods in Medical Imaging
Files, DICOM, Geometry
David Atkinson, UCL
06/04/2011 09:38
19
Computational Methods in Medical Imaging
Files, DICOM, Geometry
06/04/2011 09:38
Tutorial
• Download and unzip COMUNIX.zip PET
and CT DICOM files from David Atkinson’s
course web page.
• Look through the CT and PET datasets
datasets. A
possible way to do this is:
dirout = dir('*.dcm') ;
fns = {dirout.name}' ;
montage(fns,'DisplayRange',[])
David Atkinson, UCL
20
Computational Methods in Medical Imaging
Files, DICOM, Geometry
• Are the CT and PET slices in the same
order?
• Are the CT and PET slices parallel?
• Is
I the
th tumour
t
on the
th patient
ti t left
l ft or right?
i ht?
• Note the PET slice number containing the
approximate centre of the tumour.
• Note its anatomical location in the head
foot direction and find the CT slice
containing the tumour.
06/04/2011 09:38
• Estimate the location of the tumour centre
in patient coordinates.
• Read through the file petct_display.m,
modify to produce an image where the
tumour is at the intersection of slices:
PET montage
Geometry: Coordinates and
Vectors
y
(4,2,1)
r
x
z
r = 4xˆ + 2yˆ + zˆ
r = 4ˆi + 2ˆj + kˆ
CT montage
David Atkinson, UCL
21
Computational Methods in Medical Imaging
Files, DICOM, Geometry
06/04/2011 09:38
Projections
Scalar Product (dot product, inner product)
a = a aˆ b = b bˆ
ĵ
r.ˆj = r ˆj cos(θ )
a ⋅ b = a b cosθ
r
θ
a cosθ
b
r length of vector r
ˆj = 1 (unit vector with length 1)
θ
a
b cosθ
dot (or scalar) product is
length of projection of r onto unit vector j
[MathWorld]
Scalar product. notation
Vector Length: Norm
u = u1ˆi + u2 ˆj + u3kˆ
r = 4ˆi + 2ˆj + kˆ
v = v1ˆi + v2 ˆj + v3kˆ
r ⇒ r 2 = 4 2 + 2 2 + 12
ˆj = 0 2 + 12 + 0 2 = 1
2
3
u ⋅ v = ∑ ui vi = u1v1 + u2v2 + u3v3
i =1
Vector Cross product
• Cross product of two vectors u and v is a
third vector perpendicular to u and v.
• E.g. z axis is cross product of x and y.
u× v
Vector product
u× v
u
u = u1ˆi + u2 ˆj + u3kˆ
v = v1ˆi + v2 ˆj + v3kˆ
θ
v
u × v = ˆi (u y v z − u z v y ) + ˆj(u z v x − u x v z ) + kˆ (u x v y − u y v x )
u
u × v = ( u v sin θ )nˆ
θ
v
David Atkinson, UCL
22