performance of the GLIMPS diagnostic tool.
Bas A. de Jong S2608669
Department of nuclear medicine and molecular imaging
Universitair medisch centrum Groningen (UMCG) The Netherlands
period: 24/04/2017-17/07/2017
internship
supervisor: prof. dr. R. Boellaard, Clinical Physicist, UMCG, Department of nuclear medicine.
mentor: dr. ir. A.T.M. Willemsen, UMCG, Department of nuclear
medicine.
Abstract
The aim of this study was to analyze the impact of brain PET image quality on the performance of the GLIMPS diagnostic tool for Parkinson’s disease (PD).
GLIMPS provides a score telling how much the metabolic rate of glucose distribu- tion in a subjects brain resembles the distribution in an average patient with PD and thereby how likely the subject has PD. To study the impact of brain PET image quality on the performance of the GLIMPS diagnostic tool image data was acquired using a Hoffmann brain phantom, multiple different PET scanners, reconstruction settings and acquisition durations. Average grey and white matter activities (rep- resenting the contrast in the image) were calculated to check if they are correlated with the score. Image noise was found to have a small effect on the score after 120 seconds acquisition time. Image resolution was found to have large effect on the scores, with an offset of the scores of 1000 points upon using 10mm XYZ Gaussian smoothing. The use of different scanners also has a large effect on the scores, intro- ducing differences in scores of upto 1000 points. The use of different reconstruction settings has a moderate effect on the scores. Furthermore, the scores were found to correlate strongly with image contrast. The results of this study showed that the performance of the GLIMPS diagnostic tool depend heavily on PET image quality and harmonization or calibration to account for differences caused by systems or reconstruction settings is warranted.
Contents
1 Introduction 3
1.1 Parkinson’s disease . . . 3
1.2 Parkinson’s disease in PET imaging . . . 3
1.3 GLIMPS and PDRP subject scores . . . 4
1.4 Training the software . . . 4
1.5 Validation . . . 6
2 Methods 7 2.1 Subject position and image noise . . . 8
2.2 Multiple systems and reconstruction settings . . . 8
2.3 Grey white matter activity ratios and recovery coefficients . . . 11
3 Results 12 3.1 Subject position and image noise . . . 12
3.2 Multiple systems and reconstruction settings . . . 14
3.3 Grey white matter activity ratios and recovery coefficients . . . 16
4 Discussion 19
5 Conclusions 20
A Co-registration script 22
B Grey white mean activity script 23
1 Introduction
1.1 Parkinson’s disease
Parkinson’s disease (PD) is the second most common neuro-degenerative disease after Alzheimer disease. Neuro-degenerative disease means that the functioning of a certain type of neurons in the brain decreases as the disease progresses. Symptoms of PD are impaired motor function with characteristics like: tremor, rigidity, bradykinesia (slowness of movement) or postural instability. Furthermore, depression, dementia and tiredness can also be symptoms of the disease. When a patient suffers from multiple of these symptoms it is said that the patient suffers from parkinsonism, but this does not mean that the cause of the symptoms has to be PD. Parkinsonism can also be caused by other neuro- degenerative disorders, certain brain lesions or exposure to toxins. In PD the main cause of reduction in motor skills is attributed to the degeneration of dopaminergic neurons in the substantia nigra (see Figure 1). The substantia nigra is the part of the brain, which is important in regulating motor tasks hence the symptoms. At the moment, treatment of PD only reduces the symptoms to improve the quality of life of the patient. More research is still needed into the exact cause of the neuronal degeneration and neuropathology in PD to be able to find a cure for the disease in the future.[1]
Figure 1: Visual difference in color of substantia nigra in patient with PD compared with healthy brain.
1.2 Parkinson’s disease in PET imaging
To study the causes and progression of parkinsonism, it is important to be able to monitor not only the symptoms but also the processes in the brain of the patients. This can be
done using medical imaging techniques like functional magnetic resonance imaging (fMRI) single-photon emission computed tomography (SPECT) or positron emission tomography (PET).[2] In PET imaging a biological process specific radioactive tracer is administered to the patient. A radio-tracer emits positrons, which after slowing down annihilate with electrons from the body sending out two back-to-back 511 keV photons. These photons are detected in the PET scanner using a coincidence circuit, with which the path of the photons can be determined. Using the paths of the photons of many of these annihilation events the distribution of the tracer in the patient can be reconstructed, which can give information about the biological process. The resolution of PET scanners is limited by the movement of the positron before annihilation, scattering of the 511 keV photons in the body, a less than 180 degrees emission angle between the two photons and by properties of the scanner. In modern PET scanners the maximum resolution is just under 3 mm.[3]
Examples of biological processes which can be studied using PET are gene expression, blood flow, enzyme activity and metabolic rates. Especially the metabolic rates are interesting for our application, because in PD differences in the distribution of glucose metabolism compared to the glucose metabolism in a healthy brain are present. For example the glucose metabolism in the lentiform nucleus is reduced in a PD patient.[2]
Metabolic rates are typically imaged using 18FDG, this is an analogue of glucose, which is the main metabolic substrate for brain tissue.[3] The difference in the relative distribution of the metabolic rates of glucose in patients with PD compared to healthy controls has a typical profile or pattern, this is also true for other diseases. By imaging these metabolic rates, and comparing them with patterns found for different diseases, it is possible to perform differential diagnosis for PD.[4]
1.3 GLIMPS and PDRP subject scores
To perform differential diagnosis you need to analyze the PET images in some way to determine if the Parkinson’s disease related metabolic pattern (PDRP) is present. This can be done by a trained radiologist, but this will give you a subjective non quantitative diagnosis for the patient. Therefore, it would be beneficial, also for research purposes, to analyze the images by using trained software to recognize the pattern. The goal is by using software to retrieve a single score, which tells you quantitatively how much your FDG PET image contains the PDRP and thereby how likely the patient has got PD. For this, the GLIMPS diagnostic tool is used. The GLIMPS software can also be trained to recognize metabolic patterns of other diseases, therefore it is also possible to do differential diagnosis for multiple diseases using this tool.
1.4 Training the software
To train the software, PD patient and healthy control image data is used and some mathematical procedures are performed. This is done using the GLIMPS system. These procedures involve image registration to standard space, normalization, analysis of vari- ance and co-variance done using principal component analysis (PCA).[4]
Image registration changes the orientation and sometimes also the shape (depending on the type of registration used) of the image to fit as good as possible onto a template brain image. This way similar brain structures get the same coordinates in all images.
Examples of registration types are: rigid, affine and B-spline transformations.[6] Then a normalization step is performed to minimize the differences caused by inter-individual differences in for example global 18FDG uptake and the administered amount of 18FDG.
Normalization can also be done in multiple ways. It is possible to choose a reference region which is thought not to be affected by the disease, and normalize to the average activity in that region. In PD often the cerebellum is used for this purpose. An other method is global mean normalization. To perform the normalization the average grey matter value is calculated using a grey matter mask, subsequently the voxel values are normalized to this average value. Another option is to use the scaled subprofile model (SSM). In SSM the data is first log transformed and subsequently the log mean is subtracted. This procedure removes linear scaling factors and centers the voxel values around zero.[4] ”Log transfor- mation of the images separates essentially meaningless multiplicative scaling effects into additive components that are subsequently removed by the centering operation.”[4] After the registration and normalization procedures, a vector of the voxel values is called the subject residual profile (SRP).
After the image data is registered to standard space and normalized to remove non- disease related differences, the remaining differences can be analyzed to find the PDRP.
The most basic way to analyze the differences, is to calculate the variance between the average voxel values of the patient group and the healthy control group. Large variance in regions would mean a significant difference in metabolic rates in those regions possi- bly related to the disease. It is also possible to do a covariance analysis in which the relationship between different voxels is also taken into account. This can give you more information about the disease related pattern from the data. The downfall of covariance analysis is, when you analyze many voxels and patients, it becomes harder to appreciate the patterns in the data (because of increased dimentionality), as it is no longer possible to present it graphically. [4] A method which helps to reduce the dimensionality of the data is principal component analysis. In PCA eigenvectors (called principal components or PCs) and eigenvalues are calculated from a subject x subject covariance matrix. The PCs are arranged in an order in which the amount of variability, the PC accounts for, decreases with its number. Since the first PCs describe most of the variability in the data, a choice can be made to discard some of the later PCs ,which probably have less infor- mation about the pattern of the disease. This reduces the dimensionality of the data.[4]
The GLIMPS system uses SSM PCA for the normalization and co-variance analysis steps.
Using the inner product of the PCs and SRPs, subject scores are calculated. It is made sure that the mean subject scores of subjects with PD are higher on each PC than the mean controls. This is done by multiplying with minus one if it is not the case. On some PCs patients will score significantly higher than the controls. These PCs will be selected, because they most likely contain the disease related pattern. By performing these pro- cedures it is made sure that diseased subjects get higher scores than the controls. The
selected PCs in the end form the PDRP vector, which is used to calculate PDRP subject scores for new subjects.[4]
To calculate a subject score for a new patient. The image of the patient is processed using the same registration and normalization procedures to again get a SRP. The PDRP subject score is calculated using the inner product of SRP with the PDRP vector. PDRP subject scores for PD patients end up to be around 500 points and healthy control subject scores around -1500 points. To make interpretation of the subject score less difficult, the subject scores can be z-transformed using (1):
ZSS = SS − µSSHC σSSHC
. (1)
In which ZSS is the z-transformed subject score, SS is the PDRP subject score, µSS HC is the average PDRP subject score of the healthy control group and σSS HC is the standard deviation of the PDRP subject scores of the control group. The z-transformation sets the average score of healthy controls to zero with a standard deviation of one. When a found z-transformed score is higher than e.g. three, it deviates significantly from the mean healthy control scores, which implies that the subject might have PD.[4]
1.5 Validation
Before using the PDRP in research or for diagnosis it is important to validate the per- formance of the GLIMPS diagnostic tool. It should be checked if the calculated subject scores for new subjects indeed correlate with markers of the disease like disease duration and severity, to check if the pattern has a real meaning.[4] It should also be checked if the method delivers consistent results regardless of FDG PET image quality. Therefore, we studied the effect of different subject position, image noise, image resolution and the use of different imaging systems and reconstruction methods on the PDRP subject scores.
2 Methods
To check the impact of image quality on the performance of the GLIMPS diagnostic tool, image data was acquired using a Hoffman brain phantom (see Figure 2). The Hoffman brain phantom mimics a anatomically accurate 18FDG tracer distribution of a healthy brain.[5] Before each measurement the brain phantom was filled with about 30 kBq/ml tracer solution.
Figure 2: Hoffman brain phantom.
The images were registered using a template FDG registration mask. Normalisation was done using the SSM normalization method from SPM and a previously made PDRP.
To analyze the results, mean scores were calculated using Eq. (2):
SS = 1 n
n
X
i=1
SSi (2)
In which SS is the mean subject score, n is the number of images and SSi are the in- dividual subject scores. Also the sample variance Eq. (3) and standard deviation (SD) were calculated to get a quantitative measure of the spread of the PDRP subject scores.
s2 = 1 n − 1
n
X
i=1
(SSi− SS)2 (3)
In which s2 is the sample variance, n is the number of PDRP subject scores and SSi are the individual PDRP subject scores. The sample standard deviation is calculated as the square root of the sample variance.
To have a further investigation into the origin of the differences in the PDRP subject scores. Grey white matter activity ratios and recovery coefficients were calculated. The grey white matter activity ratio can be interpreted as the contrast in the image. To study
the relation between the grey white matter activity ratio and recovery coefficients with the subject scores the Pearson correlation coefficient was calculated using Eq. (4):
r =
n
X
i=1
(GWi− GW )(SSi− SS) s n
X
i=1
(GWi− GW )2 s n
X
i=1
(SSi− SS)2
(4)
In which r is the Pearson correlation coefficient, GWiare the individual grey white activity ratios or recovery coefficients, GW is the average grey white activity ratio or recovery coefficient, SSiare the individual PDRP subject scores and SS is the mean PDRP subject score.
2.1 Subject position and image noise
The effect of subject position and image noise was studied using ten images acquired on the Siemens Biograph using a single reconstruction setting. The brain phantom was re- positioned before measuring each image, and the images were reconstructed for acquisition times of 60, 120 and 300 seconds. An example of the images for three of the positions and the different acquisition times can be seen in Figure 3. To check the effect of image resolution on the subject scores, the images were smoothed with 0, 5, 8 and 10 mm FWHM Gaussian smoothing. The effect of smoothing on the images can be seen in Figure 4.
2.2 Multiple systems and reconstruction settings
To check the effect of the use of different scanners and reconstruction settings on the PDRP scores, image data was acquired on 7 different PET scanners from multiple hospitals. The systems we used were: two Siemens Biograph systems from the universitair medisch cen- trum Groningen (UMCG 1 and 2), Siemens Horizon from the Leids universitair medisch centrum (LUMC), GE710 from Catharina Hospital, Philips Vereos from Philips Cleve- land and a Philips Gemini and Ingenuity from VU medisch centrum (VUMC). On each scanner multiple clinically relevant reconstruction settings were used. Examples of some reconstructions of the Siemens Biograph and the Philips Vereos can be seen in Figure 5.
Figure 3: Images of three phantom positions reconstructed for 60,120 and 300 seconds acquisition time.
Figure 4: Image smoothed with 0, 5, 8 and 10mm FWHM Gaussian smoothing.
Figure 5: Six images made using different reconstruction settings on the Siemens Biograph and the Philips Vereos.
2.3 Grey white matter activity ratios and recovery coefficients
Grey white matter activity ratios and recovery coefficients were determined by first cal- culating the average grey and white matter voxel values. This was done by registering the images using a python script and Elastix [6][7][8] to a Hoffann phantom mask using rigid transformation (avoiding deformation of the images since the mask has the same shape as the phantom), to get all images properly aligned. The python script used for registration can be found in appendix A. Then average grey and white matter activity values were calculated by averaging over the voxels, which were overlapped by eroded grey and white matter masks. An image of the eroded grey and white matter masks can be seen in Figure 6. This was done with the aid of python and the nibabel package. The script for this step can be found in appendix B. The grey white matter activity ratios were calculated by dividing the average grey by the average white voxel values. Then the grey and white matter recovery coefficients were calculated by first determining the activity concentration of stock tracer solution at the start of the scan. The the recovery coefficients are calculated by dividing the average grey and white matter voxel values by the activity concentration of the stock solution. The white recovery coefficients were subtracted from the grey recovery coefficients, because the outcome of this subtraction seemed most predictive for the PDRP subject scores. This yielded the grey minus white matter activity recovery coefficients.
Figure 6: Eroded grey and white matter masks.
3 Results
3.1 Subject position and image noise
The PDRP subject scores of the re-positioned phantom data reconstructed for 60, 120 and 300 seconds can be seen in Figure 7. The corresponding mean PDRP subject scores and their standard deviations are shown in Table 1. The mean PDRP subject score remains about the same as the acquisition time is changed and the standard sample deviation decreases from 105 to 80 points, when measuring 120 seconds instead of 60 seconds and remains about 80 upon measuring 300 seconds. Figure 8 shows the PDRP subject scores for 0, 5, 8 and 10mm XYZ FWHM Gaussian smoothed images, which had a acquisition time of 120 seconds. In this figure it can be seen that, as the images are smoothed more, the PDRP subject scores get an offset of about 400, 750 and 1000 points for 5,8 and 10mm smoothing respectively. The mean PDRP subject scores and sample standard deviations for 60, 120 and 300 seconds reconstructions smoothed with 0, 5, 8 and 10mm XYZ Gaussian smoothing, can be found in Tables 2 and 3. From the mean scores, it can be seen that all PDRP subject scores for different reconstruction times get about the same offset by smoothing. From the sample standard deviations it can be seen that the spread of the scores is only significantly reduced by smoothing for the images with 60 seconds reconstruction time.
Figure 7: PDRP subject scores versus acquisition time for ten times re-positioned phantom images.
Table 1: Mean PDRP subject scores and their corresponding sample standard deviations for different acquisition times of the ten times re-positioned phantom images.
acquisition time [s] mean subject score sample standard deviation
60 -1451 105
120 -1447 80
300 -1452 81
Figure 8: PDRP subject scores versus smoothing for ten times re-positioned phantom images smoothed with 0, 5, 8 and 10 mm XYZ Gaussian smoothing
Table 2: Mean PDRP subject scores for 60,120 and 300 second acquisition time and 0, 5, 8 and 10mm XYZ Gaussian smoothing
acquisition time [s] 0mm 5mm 8mm 10mm
60 -1451 -1050 -691 -435
120 -1447 -1054 -697 -440
300 -1452 -1062 -704 -447
Table 3: Sample standard deviation of PDRP subject scores for 60,120 and 300 second acquisition time and 0, 5, 8 and 10mm XYZ Gaussian smoothing
acquisition time [s] 0mm 5mm 8mm 10mm
60 105 90 84 82
120 80 79 77 77
300 81 79 79 78
3.2 Multiple systems and reconstruction settings
The PDRP subject scores for the images acquired on different systems using different re- construction settings are displayed in Figure 9. The PDRP subject scores differ upto 1500 points. The cause of these large differences is further examined in the grey white matter activity ratio analysis. In Figure 10 the PDRP subject scores for the Siemens Biograph (UMCG1) are plotted as function of smoothing. Again it can be seen that the PDRP subject scores get an larger offset as the amount of smoothing increases. The mean and sample standard deviations of the PDRP subject scores for the different reconstructions as a function of smoothing can be seen in Table 4. The sample standard deviation without smoothing is 160 points, this is a larger standard deviation than which was found using the re-positioned phantom, were only one reconstruction method was used. Therefore, it seems that the reconstruction method used does increase the variability of the PDRP subject score. After smoothing the sample standard deviation goes down. The mean, sample standard deviation and offsets were also calculated for the GE710 and can be seen in Table 5. From these values it can be seen that smoothing has a similar effect on these subject scores, the offset is a bit smaller and the sample standard deviations are reduced upon smoothing.
Table 4: Sample mean and standard deviation of the PDRP subject scores for the 0, 5, 8 and 10mm XYZ Gaussian smoothed Siemens Biograph (UMCG1) images.
0mm 5mm 8mm 10mm
mean -1456 -1033 -673 -419
sample dev 167 157 137 123
offset 0 422 782 1037
Table 5: Sample mean and standard deviation of the PDRP subject scores for the 0, 5, 8 and 10mm XYZ Gaussian smoothed GE710 images.
0mm 5mm 8mm 10mm
mean -813 -499 -187 40
sample dev 203 165 140 122
offset 0 314 625 853
Figure 9: PDRP subject scores for images of multiple systems using different reconstruction settings.
Figure 10: PDRP subject scores versus smoothing of 0, 5, 8 and 10 mm XYZ Gaussian smoothed images of Siemens Biograph (UMCG1).
3.3 Grey white matter activity ratios and recovery coefficients
The grey white matter activity ratios were calculated for the re-positioned phantom data using multiple acquisition times and their corresponding PDRP subject scores were plot- ted versus the activity ratios in Figure 11. The mean grey white matter activity ratios and their sample standard deviations as function of smoothing can be found in Tables 6 and 7. The average grey white matter activity ratios do not change significantly as function of acquisition time, they do change significantly as function of smoothing. The latter is of course expected, as smoothing decreases the contrast in the image. The sample standard deviations of the grey white matter activity ratios also do not significantly change as function of acquisition time, but they do decrease by more than factor of 2 after 10mm smoothing. In Figure 12 the PDRP subject scores of the 120 seconds acquisition time re-positioned phantom data, which was smoothed with 0, 5, 8 and 10mm XYZ Gaussian smoothing, is plotted against their grey white matter activity ratios. The Pearson correla- tion coefficient r for this data set is -0,98, indicating an almost perfect linear relationship between the grey white matter activity ratios and PDRP subject scores under the effect of smoothing. Figure 13 shows the PDRP subject scores versus the grey white matter activity ratios for the different scanners and reconstruction settings. In this figure, it can also be seen that the two are correlated. The Pearson correlation coefficient for the data of the different scanners was calculated to be -0,65. Especially for the data of the Philips Vereos the correlation between the PDRP subject scores and the grey white activity ratio was high with a Pearson correlation coefficient of -0.991. It is also clear that the contrast in the image does not explain the whole difference in PDRP subject scores between the systems, since some subject scores of the GE710 are 600 points higher than subject scores with about the same grey white matter activity ratios from the Siemens Horizon (LUMC) scanner. The PDRP subject scores versus the grey minus white matter activity recovery coefficients of the different scanners are plotted in Figure 14. The Pearson correlation coefficient for this data was -0,79, indicating a stronger correlation between the PDRP subject scores and the grey minus white matter activity recovery coefficients compared with the PDRP subject scores and the activity ratios.
Table 6: Mean grey white matter activity ratios of re-positioned phantom data smoothed with 0, 5, 8 and 10mm XYZ Gaussian smoothing.
0mm 5mm 8mm 10mm
60s 3.46 3.03 2.60 2.34 120s 3.46 3.03 2.61 2.34 300s 3.47 3.03 2.61 2.34
Table 7: Standard deviation grey white matter activity ratios of re-positioned phantom data smoothed with 0, 5, 8 and 10mm XYZ Gaussian smoothing.
0mm 5mm 8mm 10mm
60s 0.025 0.016 0.011 0.008 120s 0.025 0.016 0.012 0.009 300s 0.020 0.015 0.011 0.008
Figure 11: PDRP subject scores versus grey white matter activity ratios of re-positioned phantom data for 60, 120 and 300 seconds aquisition time.
Figure 12: PDRP subject scores versus grey white matter activity ratios of 120 seconds acquisition time re-positioned phantom data smoothed with 0, 5, 8 and 10mm XYZ Gaussian smoothing.
Figure 13: PDRP subject scores versus grey white matter activity ratios of images from different
Figure 14: PDRP subject scores versus grey minus white activity recovery coefficients of images from different scanners using multiple reconstruction settings.
4 Discussion
The GLIMPS score or PDRP subject score has the highest value when it gives consistent and accurate results. From our study we can conclude that the use of different scanners and reconstruction settings has a large effect on the scores. To get more consistent results, it would be a good idea to either use only one type of scanner and reconstruction setting, or correct the score for the scanner and reconstruction setting used. Correcting of the scores could be done by performing a calibration measurement using the Hoffmann phantom.
Another option for correcting the scores could be to correct according to the found grey minus white activity recovery coefficients, since those values were found to be correlated the most with the PDRP subject scores for the different scanners and reconstruction methods. An explanation why the grey and white recovery coefficients are correlated with the PDRP subject scores could be, that the difference between the mean grey and white matter recovery coefficients is related with the contrast in the image. If there is a lot of contrast in the image, the metabolic pattern could be more clearly expressed in the image, hence a healthy subject would get a lower score. The contrast is reduced upon smoothing, this also reduces how clearly the metabolic pattern is expressed, and the PDRP subject score would be higher. For further studies into the impact of brain PET image quality on the performance of the GLIMPS diagnostic tool, it could be informative to perform the same experiments we did, using a brain phantom, which can mimic the
18FDG concentration distribution (pattern) found in diseased subjects. This would allow to also study how diseased subject scores depend on the FDG PET image quality.
5 Conclusions
It is beneficial to have a method of diagnosing PD and have a quantitative measure of how likely the disease is present in a patient. For the PDRP subject score to have most value it is important that the method gives consistent results. From the analysis of our data, it can be concluded that re-positioning of the phantom results in a standard deviation of the PDRP subject scores of 80 points. Having a longer data acquisition time than 120 seconds does not seem to increase the precision or accuracy of the PDRP subject scores. The use of different scanner systems and reconstruction settings does have a large effect on the found PDRP scores. Smoothing of the data can reduce these differences, but also introduces a large offset to the found PDRP subject scores. Smoothing also reduces the grey white matter activity ratios, which are highly correlated with the found PDRP subject scores. Grey white matter activity ratios do not explain the whole difference in the found PDRP subject scores between scanner systems, as PDRP subject scores of different scanners can have the same grey white matter activity ratios, while there is a difference of 600 points in PDRP subject scores. A stronger correlation with the PDRP subject scores was found for the grey minus white activity recovery coefficients compared with the activity ratios.
Acknowledgements
I would like to thank Ronald Boellaard, for giving me this subject for my internship and his guidance during the project. Furthermore, I would like to thank Elizabeth Pfaehler who was my daily super visor and Rosalie Kogan who were always open for questions.
References
[1] W. Dauer, S. Przedborski, Parkinson’s Disease: Mechanisms and Models, Neuron, Volume 39 issue 6, september 2003, Pages 889–909
[2] DJ Brooks Imaging approaches to Parkinson disease, Journal of medicine, April 2010, 51(4), pages 596-609.
[3] Basu, S Hess, S and Nielsen, P E Olsen, B B Inglev, S Høilund-Carlsen, P F The Basic Principles of FDG-PET/CT Imaging., PET clinics, October 2014, 9(4):355-70, [4] P.G. Spetsieris, D.Eidelberg Scaled Subprofile Modeling of Resting State Imaging Data in Parkinson’s Disease: Methodological Issues, Neuroimage, February 2011, 54(4), pages 2899-2914.
[5] Hoffman 3-D Brain Phantom, http://www.biodex.com/nuclear- medicine/products/phantoms/hoffman-3-d-brain-phantom, Biodex.
[6] S. Klein, M. Staring Elastix the manual,
http://elastix.isi.uu.nl/download/elastixmanualv4.8.pdf, September2015.
[7] S. Klein, M. Staring, K. Murphy, M.A. Viergever, J.P.W. Pluim, ”]elastix: a tool- box for intensity based medical image registration, IEEE Transactions on Medical Imaging, January 2010, vol. 29, no. 1, pages 196 - 205.
[8] D.P. Shamonin, E.E. Bron, B.P.F. Lelieveldt, M. Smits, S. Klein and M. Staring, Fast Parallel Image Registration on CPU and GPU for Diagnostic Classification of Alzheimer’s Disease, Frontiers in Neuroinformatics, January 2014, vol. 7, no. 50, pp.
1-15.
A Co-registration script
1 # −∗− c o d i n g : u t f −8 −∗−
2 3 ”””
4 T h i s s c r i p t co−r e g i s t r a t e s n i f t y PET FDG B r a i n i m a g e s t o a mask u s i n g e l a s t i x .
5 Arguments t o t h e s c r i p t a r e f i l e s and f o l d e r s which s h o u l d be s e l e c t e d from
6 a f i l e e x p l o r e r window ( o p e n s b e h i n d python e d i t o r a f t e r you run t h e s c r i p t
! ! ! ) .
7 Arguments o f t h e f u n c t i o n a r e : F o l d e r c o n t a i n i n g n i f t y image f i l e s ,
8 Mask f i l e ,
9 Parameter f i l e
10 ”””
11 import o s #Import modules
12 import s u b p r o c e s s
13 from t k i n t e r import∗
14 from t k i n t e r import Tk , f i l e d i a l o g
15
16 ””” s e l e c t arguments f o r ’ r e g i s t r a t i o n ’ f u n c t i o n from f i l e e x p l o r e r window
”””
17 Tk ( ) . withdraw ( ) # we don ’ t want a f u l l GUI , s o keep t h e r o o t window from a p p e a r i n g
18 i m a g e F o l d e r = f i l e d i a l o g . a s k d i r e c t o r y ( t i t l e = ’ s e l e c t f o l d e r c o n t a i n i n g n i i f i l e s ’ ) # show an ”Open” d i a l o g box and r e t u r n t h e path t o t h e s e l e c t e d d i r e c t o r y
19 #p r i n t ( i m a g e F o l d e r )
20
21 Tk ( ) . withdraw ( )
22 m a s k f i l e = f i l e d i a l o g . a s k o p e n f i l e n a m e ( t i t l e = ’ s e l e c t co−r e g i s t r a t i o n mask f i l e ’ )
23 #p r i n t ( m a s k f i l e )
24
25 Tk ( ) . withdraw ( )
26 p a r a m e t e r F i l e = f i l e d i a l o g . a s k o p e n f i l e n a m e ( t i t l e = ’ s e l e c t p a r a m e t e r f i l e ’ )
27 #p r i n t ( p a r a m e t e r F i l e )
28 29
30 ””” co−r e g i s t r a t i o n o f i m a g e s t o mask f i l e ”””
31 o s . c h d i r ( ”C: / TemporaryUserSpaceBdJ / python ” )
32 def c o r e g i s t r a t i o n ( i m a g e F o l d e r , m a s k f i l e , p a r a m e t e r F i l e ) :
33 c o R e g i s t r a t e d I m a g e F o l d e r= i m a g e F o l d e r + ’ c o −r e g i s t r a t e d ’ #d e f i n e f o l d e r which c o n t a i n s c o r e g i s t r a t e d i m a g e s & r e g i s t r a t i o n i n f o
34 o s . mkdir ( c o R e g i s t r a t e d I m a g e F o l d e r ) #c r e a t e f o l d e r which c o n t a i n s o u t p u t
35 print ( ” co−r e g i s t r a t e d i m a g e s i n : ” , c o R e g i s t r a t e d I m a g e F o l d e r )
36 # c o u n t V a l u e=0
37
38 f o r image in o s . l i s t d i r ( i m a g e F o l d e r ) : #C r e a t e l i s t o f image f i l e s & l o o p o v e r Images
39 # c o u n t V a l u e+=1
40 imagePath=( i m a g e F o l d e r + ’ / ’+image )
i f imagePath . e n d s w i t h ( ” n i i ” ) : #s e l e c t o n l y n i f t y f i l e s t o be
42 outputFolderName = c o R e g i s t r a t e d I m a g e F o l d e r + ” / r e g i n f o ”+ image [ : − 4 ] #d e f i n e o u t p u t f o l d e r n a m e a s i n p u t f o r e l a s t i x
43 o s . mkdir ( outputFolderName ) # c r e a t e s e p a r a t e r e g i s t r a t i o n i n f o f o l d e r s f o r i m a g e s
44
45 command= ( ’ e l a s t i x ’ , ’−f ’ , m a s k f i l e , ’−m ’ , imagePath ,
46 ’−o u t ’ , outputFolderName , ’−p ’ , p a r a m e t e r F i l e ) #d e f i n e command which has t o be e n t e r e d i n cmd
47 s u b p r o c e s s . c a l l ( command )
48 # p r i n t ( s t r ( command ) )
49
50 #rename f i l e p ro du ce d by e l a s t i x t o o r i g i n a l f i l e name w i t h ’R ’ p r e f i x &
put i t i n o u t p u t f o l d e r ”
51 oldPathAndFilename=outputFolderName + ” / r e s u l t . 0 . n i i ”
52 newPathAndFilename= c o R e g i s t r a t e d I m a g e F o l d e r + ” /R” + image
53 o s . rename ( oldPathAndFilename , newPathAndFilename )
54 continue
55 e l s e :
56 continue
57 c o r e g i s t r a t i o n ( i m a g e F o l d e r , m a s k f i l e , p a r a m e t e r F i l e )
58 59 ”””
60 C r e a t e d on Tue May 23 1 1 : 2 2 : 5 1 2017
61
62 @author : Bas de Jong
63 ”””
Listing 1: Python example
B Grey white mean activity script
1 # −∗− c o d i n g : u t f −8 −∗−
2 3 ”””
4 T h i s s c r i p t c a l c u l a t e s a v a r a g e g r e y and w h i t e m a t t e r a c t i v i t i e s f o r phantom n i f t y images , which a r e a l r e a d y co−r e g i s t r a t e d t o a Hofmann b i n a r y mask .
5 The s c r i p t a v a r a g e s t h e v o x e l v a l u e s o f t h e v o x e l s which a r e o v e r l a p e d by
6 e r o d e d Grey and White m a t t e r masks .
7 ! ! ! F o l d e r and f i l e s e l e c t i o n windows may open b e h i n d phyton e d i t o r ! ! !
8 ”””
9 import o s
10 import n i b a b e l a s n i b
11 import numpy a s np
12 from t k i n t e r import Tk , f i l e d i a l o g
13
14 ” s e l e c t arguments f o r ’ g r e y W h i t e V a l u e s ’ f u n c t i o n from a f i l e e x p l o r e r window ”
15 Tk ( ) . withdraw ( ) # we don ’ t want a f u l l GUI , s o keep t h e r o o t window from a p p e a r i n g
16 i m a g e F o l d e r = f i l e d i a l o g . a s k d i r e c t o r y ( t i t l e = ’ s e l e c t f o l d e r c o n t a i n i n g co−
r e g i s t r a t e d n i i f i l e s ’ ) # show an ”Open” d i a l o g box and r e t u r n t h e path t o t h e s e l e c t e d d i r e c t o r y
17
18 Tk ( ) . withdraw ( )
19 g r e y M a s k F i l e = f i l e d i a l o g . a s k o p e n f i l e n a m e ( t i t l e = ’ s e l e c t e r o d e d g r e y m a t t e r mask f i l e ’ )
20
21 Tk ( ) . withdraw ( )
22 w h i t e M a s k F i l e = f i l e d i a l o g . a s k o p e n f i l e n a m e ( t i t l e = ’ s e l e c t e r o d e d w h i t e m a t t e r mask f i l e ’ )
23 ” d e f i n e f u n c t i o n which c a l c u l a t e r a v a r a g e g r e y and w h i t e m a t t e r v a l u e s u s i n g e r o d e d masks ”
24
25 def g r e y W h i t e V a l u e s ( i m a g e F o l d e r , g r e y M a s k F i l e , w h i t e M a s k F i l e ) :
26 greyMask=n i b . l o a d ( g r e y M a s k F i l e ) #u s e n i b a b e l t o l o a d n i f t y mask f i l e
27 g r e y M a s k a r r a y=greyMask . g e t d a t a ( ) #c r e a t e g r e y mask a r r a y
28 whiteMask=n i b . l o a d ( w h i t e M a s k F i l e )
29 w h i t e M a s k a r r a y=whiteMask . g e t d a t a ( ) #c r e a t e w h i t e mask a r r a y
30 t a b l e =[ ” image name ” , ’ g r e y m a t t e r mean ’ , ’ w h i t e m a t t e r mean ’ , ’ g r e y / w h i t e r a t i o ’ ] #d e f i n e t op p a r t o f t a b l e
31 c o u n t V a l u e=0
32 f o r image in o s . l i s t d i r ( i m a g e F o l d e r ) : #C r e a t e l i s t o f image f i l e s &
l o o p o v e r i m a g e s
33 imagePath=( i m a g e F o l d e r + ’ / ’ +image )
34 i f imagePath . e n d s w i t h ( ” . n i i ” ) :
35 img=n i b . l o a d ( imagePath )
36 i m g a r r a y=img . g e t d a t a ( ) #c r e a t e image a r r a y
37 g r e y M a t t e r=i m g a r r a y [ ( g r e y M a s k a r r a y >0) ] #s e l e c t p a r t o f image f i l e o v e r l a p p e d by greyMask
38 greyMatterMean=np . mean ( g r e y M a t t e r )
39 w h i t e M a t t e r=i m g a r r a y [ ( w h i t e M a s k a r r a y >0) ] #s e l e c t p a r t o f image f i l e o v e r l a p p e d by whiteMask
40 whiteMatterMean=np . mean ( w h i t e M a t t e r )
41 g r e y W h i t e R a t i o=greyMatterMean / whiteMatterMean
42 v a l u e s =[ image , s t r ( greyMatterMean ) , s t r ( whiteMatterMean ) , s t r ( g r e y W h i t e R a t i o ) ] #a r r a n g e i n t r e s t i n g v a l u e s i n v e c t o r
43 t a b l e= t a b l e . a d d ( v a l u e s ) #add v a l u e s t o ” t a b l e ” f o r e a c h l o o p ( c r e a t e s v e c t o r )
44 c o u n t V a l u e+=1
45 print ( ” p r o c e s s e d image ” , c o u n t V a l u e )
46 continue
47 e l s e :
48 continue
49 l e n g t h T a b l e= len ( t a b l e )
50 t a b l e = np . r e s h a p e ( t a b l e , ( l e n g t h T a b l e / / 4 , 4 ) ) #r e s h a p e t a b l e t o form a a r r a y w i t h 4 c o l l u m n s and l e n g t h T a b l e /4 rows
51 print ( ” d a t a i n m a t r i x : ” , t a b l e )
52 t x t O u t p u t F i l e= i m a g e F o l d e r + ’ / g r e y W h i t e v a l u e s . t x t ’ #s e t o u t p u t f i l e name t o end up i n same f o l d e r a s image f i l e s
53 print ( ” o u t p u t f i l e l o c a t i o n : ” , t x t O u t p u t F i l e )
54 np . s a v e t x t ( t x t O u t p u t F i l e , t a b l e , fmt= ’%s ’ , n e w l i n e= ’ \ r \n ’ , d e l i m i t e r = ’ \ t ’ ) # s a v e d a t a i n . t x t o u t p u t f i l e
55
56 g r e y W h i t e V a l u e s ( i m a g e F o l d e r , g r e y M a s k F i l e , w h i t e M a s k F i l e )
58 59 60 ”””
61 C r e a t e d on F r i Jun 2 1 4 : 3 5 : 1 6 2017
62
63 @author : GLIMPS project
64 ”””
Listing 2: Python example
