COST B11 Working Group 3 - Texture formation
Chairman: Prof Dr Lothar R. Schad, Heidelberg
Protocoller: Anne M. Fenstad and Arvid Lundervold, Bergen
The WG 3 meeting was held on Friday 5th November from
9:00-10:30 AM. There were three talks:
- Standardized MR Pulse Sequences for Test Object
Imaging: What has to be considered? (slides)
Prof. L. R. Schad, German Cancer Research Center, Heidelberg.
- Stability, Transferability and Normalization of
Texture Parameters. (slides)
Prof. I. Zuna, German Cancer Research Center, Heidelberg and Czech Technical University in Prague.
- How does the signal-to-noise ratio influence texture
Dr. A. Lundervold and A. M. Fenstad, Department of
Physiology, University of Bergen.
Standardized MR Pulse Sequences for Test Object
Imaging: What has to be Considered?
The purpose is to agree on a standardized MR pulse sequence
that one can perform on the same test object using different
MR scanners and obtain similar texture measure results.
Five different groups of
MR parameters must be considered. These are related to: contrast,
slice, coil, resolution/SNR, and reconstruction,
There will have to be an agreement on the setting
of these parameters.
A conventional spin-echo (SE) sequence is suggested as a first choice.
This because of
reduced artifact sensitivity, availability and
the possibility to use T1, T2 and proton contrasts in SE acquisitions. The
disadvantages are that the image contrast depends on quality of
rf-pulses and a slice cross-talk.
Regarding slice parameters, one has to decide on
thickness and position of the slices, and the excitation
order - linear (ascending/descending) or
interleaved slice order.
For the resolution parameters one must consider field-of-view (FOV),
the reconstructed matrix, measured k-lines, rectangular
FOV ratio, read/phase oversampling and number of acquisitions, aiming at
a good signal-to-noise ratio.
On the Siemens Vision there are two possibilities for filter settings in
reconstruction of the
image. One can use raw data filtering (Hamming/Hanning, oversampling)
or image filtering (mean, median). Some of this could be "hidden" in
the reconstruction process and might vary between scanners from
different MR manufacturers.
Basically, there are two different coils: transmit and receive.
Coil parameters will influence RF-excitation profile and sensitivity. For
receive only coils (surface coils and phased array coils), one must
consider mechanisms for inhomogeneity correction and image combination.
After considering all these parameters the proposal
for a MR-protocol will be:
- Spin Echo Sequence
- TR/TE=200-400 ms / 5-15 ms
- Matrix: 128^2 - 512^2
- FOV : 150-300 mm
- SL : 3-5 mm
- NEX : 1-8
- NO Filter
- Head Coil
Stability, Transferability and Normalization of
The talk gave a description of texture formation
in medical ultrasound images.
An overview of texture parameters and texture analysis
from previous work on ultrasound was summerized:
One important issue was the stability of the parameter estimates. The
stability were very dependent on ROI-size. If the
ROIs were greater than 800 points there seems to be no problem.
If the ROI was between 400 and 800 points
there was an instability effect for some parameters, and for ROI less than 400
points there were no valid evaluatin possible.
An other important message was that there should
be an equal distribution of ROI-sizes over classes.
These experiences (over many years) with texture in medical ultrasound
could also be applicable to MRI and should be considered in following work.
- first-order greylevel statistics (mean, variance, skewness,
curtosis and percentiles)
- first-order gradient statistics (mean gradient, variance of
gradients and relative frequency of edge elements)
- co-occurrence matrix (contrast, homogeneity, entropy and
- run-length matrix (run percentages, long-run emphasis,
greylevel distribution and runlength distribution)
How does the signal-to-noise ratio influence texture
The measured signal in MRI (complex-valued k-space) can be identified as
the true signal added a Gaussian noise term.
The noise variance of this thermal noise
is a sum of independent stochastic processes related to the body, the coil,
and the electronics.
Definitions of signal-to-noise ratio (SNR) and contrast-to-noise ratio
(CNR) were given together with
methods for parameter estimation from observed data.
In the noise/texture experiments three groups of textural
features were used: histogram-based, gradient-based and
co-occurrence matrix-derived parameters (i.e. mean,
variance, mean absolute gradient, variance of absolute
gradient, angular second moment, contrast and correlation).
All parameters were calculated within user-defined regions of interest
(ROIs) using the MaZda program (v.2.11) developed within the
COST B11 by the Lodz group. In addition,
Matlab was used for
analysis and graphics, XV
was used for color-table editing, and
(i.e. xregion) was used for detailed drawing of ROIs.
Three datasets were used in the analysis: the "straws data" from Heidelberg
(Imaging of Test Objects), raw (k-space) data from a SE head acquisition, and data from a breast
Details about the acquisition parameters and noise characteristics
are given in the slide
A few significant findings and observations from this study are:
Results from this study are planned to be presented in a
separate paper. A short abstract is also submitted to ISMRM 2000 (pdf).
- Texture measures can be sensitive to spatial sampling (choice of ROI) of
a given material or tissue type.
measures can be sensitive to ROI size.
- Texture measures is sensitive to thermal
noise and can show substantial variation even at a fixed noise
- Texture measures can exhibit a nonlinear relationship to the amount
After each talk there were given time for discussion and
Back to COST B11 homepage