Quantitation of Magnetic Resonance Image Texture
BRUSSELS WORKING GROUP - 25th - 28th June, 1998
A Review of Texture Methods and Applications in MRI
Present: Dr R A Lerski, Dundee Dr Y Rolland, Rennes
Prof Dr L R Schad, Heidelberg Dr M Bock, Heidelberg
Prof Dr A Materka, Lódz Dr M Strzelecki, Lódz
Mr P Szczypinski, Lódz Prof A Spisni, Parma
Prof Dr R Dommisse, Antwerp Dr P Ring, Copenhagen
Dr O Yu, Strasbourg Ms S Allein, Brussels
Dr H Stødkilde-Jørgensen, Aarhus Ass Prof A Santos, Madrid
Mr N Malpica, Madrid Prof R Muller, Mons
Mr A Amorison, Mons Mr H Langenberger, Vienna
Ass Prof A Lundervold, Bergen Mrs A Barclay, Dundee
Thursday 25th June, 1998
1) Welcome and Introductions
The proceedings commenced with R Lerski (Chairman, Cost B11) welcoming
everyone to Brussels for the first Working Group meeting of COST Action
B11, for which there is funding for the next four years (started March 1998).
A presentation was given by R Lerski to explain the methods by which the
COST programme functions, along with some of the bureaucracy involved.
In particular, reimbursement to delegates appears to take around 3 months,
but Dr Lerski will write to see if this can be “speeded up”.
R Lerski has applied for a subsidy to set up a Web-page describing the
Action, but so far, this has been blocked by the Commission. It is hoped
that this is temporary.
2) Presentation by Dr Y Rolland - The Clinical Value of Texture Analysis
Y Rolland started by stating that the questions to be answered were:
“The problems yesterday - can we perform texture analysis”?
“The problems today - how to perform texture analysis”?
He commented that it was technically easy to carry out texture analysis but
often the exact purpose of the experiment is not clear. It is easy to
demonstrate results, but not so easy to compare the clinical value. He also
proposed that texture analysis could be added to information already
available to Radiologists in order to find a better diagnosis.
R Muller asked in what specific cases it may be helpful. Would it be help for
the image discrimination and help for the diagnosis, as they are two
different things? In some pathologies there is no real requirement for
Y Rolland stated that it was difficult to know how to use it:
a) Do it manually, e.g., as a radiologist with visual interpretation?
b) Do it with computer, i.e., as we propose with NMRWin.
He recommended writing exactly what one is reporting - try to standardise.
L Schad - aim from beginning - is it better to have tissue characterisation
now we have texture? He said that he hoped it would be possible to
differentiate between normal and abnormal tissue, but asked the questions:
a) How we can normalise?
b) How we can standardise?
H Stødkilde-Jørgensen commented that it would probably be necessary to
try to combine texture analysis with other information.
R Lerski stated that, in his experience of texture analysis in MRI, it was
usual to experience different results on the same machine and also on
different machines. He also mentioned that, so far, the technique and
protocol have not been sufficiently developed and that some means of
standardisation or processing would be necessary.
M Bock thought it necessary to have specific information as to exactly what
we would be looking for in pathological terms so that we can predict which
parameter would be good, and which ones would not be.
3) Presentation by R Lerski - Texture Analysis in Ultrasound and X-ray
A review was given of the results achieved by texture analysis in X-ray and
in ultrasound. These date back to 1976.
R Muller asked - “What was the end image”. If the liver was cirrhotic or
normal - can you see any difference on CT? Y Rolland replied that there
were not many differences.
Y Rolland also stated that it wasn’t clear in the case of ultrasound what the
pixel meant as the resolution is too poor, therefore it is not obvious what
you are looking at or measuring.
4) Presentation by Prof Dr L R Schad - Texture analysis in MRI and status
L Schad gave a full review of the usefulness of texture analysis in MRI and
commented that he would want to find out where the limits are and what
are the best parameters.
A Lundervold queried whether the results of all echoes or only a single echo
L Schad replied that 4 images were usually used:
T1 parameter image mono exp
T2 parameter image mono exp
T1 dependent image
T2 dependent image
Most radiologists use all or some of these images for diagnosis
Friday 26th June, 1998
1) Presentation by Dr R A Lerski - Results and Methods
An introduction was given of this session, viz., the methods of texture
analysis, multiparameter evaluation and test object results from the Biomed
2) Presentation by Prof Dr A Materka - Texture Analysis Methods
A full report was distributed summarising an initial summary of the
literature in texture analysis. A detailed and excellent review was given of
the various methods available and an example of an application in skin
pathology was presented.
Y Rolland asked about the work done on skin pathology - is it worth doing
it? A Materka’s reply was that, yes, his clinical collaborators think so, and
that is the important issue.
Y Rolland commented that a method for identifying discriminating
parameters can always be found.
3) Presentation by Ass Prof A Santos - Multiparameter evaluation methods
A detailed review was given of the various methods available to evaluate the
texture parameter results, e.g., discriminant analysis, principal component
analysis (PCA), neural networks.
A Materka asked if PCA can be used for image classification. The reply was
that it is useful if there are a lot of variables and also a useful tool for
4) Presentation by Dr R A Lerski - Test Object Construction and Results
R Lerski explained the construction of the Test Objects, and subsequent
results, from the previous BIOMED Concerted Action.
R Lerski continued by asking the following questions:
a) Why is the study on individual machines so different?
b) What is it that changes the parameters?
c) Is it signal-to-noise?
d) Is it different field strengths?
M Bock asked about the difference between Siemens and Philips machines.
S Allein commented that she thought the different fields strengths were a very
important factor due to possible susceptibility difficulties with the test objects.
A Lundervold wondered if it may be the calibration of the reconstruction
R Lerski thought it may be the same argument as with T1 and T2. That is,
it is important to very carefully determine how to carry out the
measurements well before deciding on their value.
A Lundervold suggested that it may not be possible to measure at different
centres, but it would nevertheless be good to get the same separation of
textures on different machines.
H Stødkilde-Jørgensen asked whether the measurements had been
completed on the same scanners, to which R Lerski replied that they were,
however, the variation wasn’t as much. The same test tubes were also sent
round the various centres.
R Lerski suggested it may be a good idea to scan the tubes very extensively
for 3/4 days on one machine and then spend time analysing the data.
L Schad commented that he wasn’t really sure if the same parameters were
used in the image reconstruction by all manufacturers or in all machines,
and suggested it may be better to do the reconstruction using Fourier
transform ourselves to determine that the settings were actually the same.
R Lerski stated that maybe we should have distributed the actual sequence
to all centres.
Y Rolland wondered why MR was used instead of CT, which has less
parameters and may be simpler to understand. L Schad replied that MR
had a better soft tissue contrast. Nevertheless, Y Rolland suggested trying
CT with foam only.
A Lundervold thought maybe another sequence would be more relevant, e.g.
diffusion. Wouldn’t the biological tissue be more obviously different? He
asked if variability mattered - many other parameters that can be used. R
Lerski stated that trabecular bone had been used for same experiments but
didn’t get very good separation.
L Schad suggested that if we changed to other parameters (e.g. diffusion,
perfusion) he doesn’t think different lesions would be separated. We would
have to first find out which machine parameters influence our outcome.
A Santos asked if particular texture parameters are machine dependent. R
Lerski thought we should look at this as there is enough existing data to
carry this out, and it would give a better indication of problems before
carrying out a new trial.
A Santos has access to a 4.7 Tesla machine and R Lerski to a 7 Tesla
R Lerski - Assuming that texture is same as tube is moved across - maybe
more than 6 ROIs should be taken, and that maybe something better than
foam tubes could be used.
R Lerski commented that trabecular bone is not easy to get in lower age
groups and that it is a problem to remove the tissue out of the bone. Also,
the texture is not the same over the whole area of bone.
L Schad suggested the possibility of sending someone round each centre
involved to measure the test objects or tubes, and asked if it could be done
as a short term mission. This would also ensure that all parameters would
definitely be the same.
1) Clinical Problem
Technical point of view - organ with not much image artefacts is required.
White matter diseases, i.e. MS, but the problem is that the spinal
cord is not studied
Brain Lesions (tumour oedema)
There is a large amount of existing data from brain and foam tubes.
1) Markov - Texture Foam (simulation)
2) Re-analysis of foams and estimate SNR in gel
Have been measured already: Different FOV
Different Slice Width
- Short term mission from Lódz to Heidelberg.
- To consider further developments
- New features e.g., ROI pixel storage
‘Other’ texture methods
Source of NMRWin could possibly be distributed but an alternative
might be the following:
- NMRWin Interface to MR images etc
Analysis in ? MATLAB/IDL
- Store ROI data (pixel values)
Groups with access to MATLAB: Lódz
Groups with access to IDL: Copenhagen
Need a full specification for ROI pixel storage.
UNSCRAMBLER (PC or UNIX) }
R (S/Plus) } Commercial
SPSS } Packages
Data on Excel (macro) - R Lerski to send to read log files.
Foams ? - Reticulated
Biological? - Trabecular bone
Existing trial images for downloading - R Lerski to send out.
Foams - composite object, sepharose, HPLC - 300 m ? - Perkin Elmer, gels
- polysaccharide, fibres, membrane
The present situation is as shown:
Histogram equalisation could be attempted
Could try a sum of parameters from brain and foam tubes to normalise.
Literature (time for mission)
Texture model (big task)
Internal exchange of partners
White matter brain
Dynamic effects (texture in time domain and spatial domain)
Same ROI in a series of slices
Breast MRI - contrast enhanced
Saturday 28th June, 1998
1) Functional MRI
Existing data - texture analysis
Aarhus, Bergen, Strasbourg and Mons
2) Library search - Lódz
3) Markov - Lódz
4) Variation of texture in reticulated foams with FOV, NSA - Dundee would
perform a full analysis
5) New Data on reticulated foam - SNR constant - Heidelberg (foams from
Lódz Heidelberg (short term mission)
2) Data Collection
Foam Images (from Dundee)
Lódz - Markov modelling
FTP of existing files - make available from Dundee
FOV, RES, NSA, Slice width
Analyse in Dundee
3) New Measurements
Foam tubes (4)
SNR Constant Heidelberg
NMRWin stores ROI - Lódz Heidelberg
Dundee will talk to London about texture routines in MATLAB
Madrid, Bergen and Lódz
DICOM reading - Madrid will investigate
Unscrambler - details to all (from Dundee)
Macro from NMRWin to Excel (from Dundee)
Madrid to look at:
Parameters from foam (Dundee will send)
6) Foam X-ray CT and Miscellaneous
Rennes will scan the reticulated foams by x-ray CT
Brussels} Wrist machine to investigate
Literature clinical ‘hot topics’ - Rennes
7) Test Objects
Antwerp will think about alternatives
Sepharose - HPLC
Parma will investigate
Gauze ? - Dundee
Embolisation beads - Rennes
Foam - T1, T2 to be changed - Dundee
T1 100 at 1.5 Tesla
T2 700 “
8) Slice Thickness
Does it vary between machines or with sequence protocol?
Plots of the variation of texture parameters with slice width are needed.
Slice width variation
0.5 1.5 2.5 3.5 4.5
Heidelberg will make the measurements, Dundee the analysis.
9) Multicentre Trial
It might be possible to look at the same ‘Siemens’ sequence
Bruker 3 Tesla
Resolution 100 m
This centre is involved in a study of the contribution of texture analysis in
the evaluation of cryptogenic temporal lobe epilepsy. Texture analysis is
being applied to standard brain MR imaging of epilepsy patients with
studies of the regions of the ipsolateral and controlateral hippocampus and
the region of amydata. Texture analysis should be used for the detection of
mesial temporary slcerosis and for the assessment of apparently normal
DATE AND VENUE OF NEXT MEETING
Thursday 19th Saturday 21st/Sunday 22nd November, Strathclyde Graduate
Business School, Glasgow, Scotland. Full details to be circulated in September.
- 10 -