Physica Medica: European Journal of Medical Physics - Volume 32 - Supplement 1 - February 2016

137 / 146

fMRI, physiological noise, due mainly to heartbeat and respiration, affects
and compromises the data and limits the analysis. RETROICO is a very
standardized method for physiological noise correction but it suffers from
the drawback of needing simultaneous respiration and heartbeat record-
ing. Conversely, CompCo is a physiological noise correction technique
with no need of such recording and that uses PCA to find nuisance com-
ponents to be regressed out in a general linear model.
In this work it was investigated whether using kernel PCA, a non-linear para-
metric form of PCA, can give for some parameter values better noise removal
performances than PCA.
Materials and Methods: The 2 methods, PCA and Gaussian kernel kPCA,
were compared by monitoring the cross-correlation between the signals
corrected via PCA or kPCA and those corrected via RETROICOR. The anal-
ysis was carried out on specific gray matter areas. The resting state fMRI
data was acquired with a 3 T Simens Skyra scanner on healthy volun-
teers. Cardiac and respiration cycles were respectively recorded with a fiber-
optic finger pulse oximeter and with a belt stretch transducer. For data
analysis, home-made MATLAB codes were used while 3D Slicer 4.4.0 was
used to select specific brain areas.
Results: The results showed that for particular values of the kernel width
the kPCA performs better in reproducing the BOLD signal and thus cor-
recting the signal from the physiological noise.
Conclusion: This preliminary data suggest that kPCA, when the kernel is
adequately chosen, can outperform the more standard PCA providing a better
solution for correction with no physiological recording.
References
[1] Glover et al., MRM.
[2] Behzadi et al., NeuroImage.
E.447
MRI QUANTIFICATION OF NON-GAUSSIAN WATER DIFFUSION IN WHITE
AND GRAY MATTERS: AN IN-VIVO DKI STUDY
L. Orsingher L. Nocett S. Piccinin G. Crisi a Servizio di Fisica Medica,
IRCCS-ASMN, Reggio-Emilia, Italy; b Servizio di Fisica Medica, Policlinico,
Modena, Italy; c Neuroradiologia, AOU, Parma, Italy
Introduction: DKI has a great potential in detection of tissue microstruc-
ture change but its clinical implementation is dampened by several practical
issues. We aim to investigate the MD, MK and FA parameters variability
due to different post-processing algorithms and different acquisition schemes
with clinically feasible times.
Material and Method: 5 healthy volunteers were imaged on a 3 T MR
system. DKI acquisitions were performed with a TRSE sequence. All images
were inspected for motion and corrected. The diffusion tensor was calcu-
lated with different b-values sets with variable bmax from 800 to 3000 s/
mm2 and with and without the K term (indicated in the following with mono
and kurt suffixes, respectively). Histogram analysis of MD, FA and MK was
extracted for WM, GM and CSF. OLS and WLLS algorithms were com-
pared against the gold standard, NLS, which was previously validated in
phantoms.
Results: WLLS results of MD, MK and FA are equivalent of those of NLS but
with a great spare of time. MD and MK histogram distributions changes
with the b-range, their median values decreasing with bmax. MD kurt is
systematically higher than MDmono but their %difference is bmax depen-
dent, ranging from 11% with bmax = 3000 s/mm2 to 23% with bmax = 800 s/
mm. Decreasing bmax from 3000 to 800 s/mm2, in GM there is an increase
of 4% and 15% in MDmono and MDkurt values respectively. The opposite
occurs to WM, with an increase of 37% and 14% in MDmono and MDkurt.
Increasing bmax, MK decreases. The term MK*MD2 is constant.
Conclusion: We found that WLLS is the best compromise between time
and accuracy. MD and MK are highly dependent on b range. The trend of
MK, if considered as a non-Gaussian behavior metric, is counter-intuitive.
However, it can be explained with a global framework approach based on
MK*MD2 parameter. We therefore suggest to analyze kurtosis results in terms
of the combined parameter MK*MD2 in addition to just taking into account
MD and MK separately.
E.448
DKI AND PSEUDO-KURTOSIS IN PHANTOMS
L. Orsingher L. Nocetti G. Crisi a Servizio Fisica Medica, IRCCS-ASMN,
Reggio Emilia, Italy; b Servizio Fisica Sanitaria, Policlinico, Modena, Italy;
c Neuroradiologia, AOU, Parma, Italy
Introduction: DKI acquisitions are affected by low SNR and Rician noise.
The present study aimed for: (1) the development of a phantom for quality
control protocol (2) the investigation of DKI parameters variability (3) the
optimization of the acquisition and post-processing workflow for clinical
application.
Materials and Methods: The experiments were conducted on a 3 T MR
systemwith a TRSE diffusion sequence. Isotropic aqueous test-solutions with
sucrose concentrations up to 30% w/w were prepared and fully character-
ized. Asparagus stems were used as biological anisotropic phantom. MD,
MK and FA values were calculated with OLS, WLLS and NLS algorithms, dif-
ferent b-values sets with variable bmax from 800 to 3000 s/mm2 and with
and without the K term. The NLS gold standard algorithm were validated
against reference D value, measured apart with a DOSY-2D sequence, and
microscopy images for free water and asparagus, respectively.
Results: The 25% and 30% w/w concentrations match the D of WM and GM
respectively. In aqueous solutions, a combined influence of signal at b = 0,
SNR, noise, b-range, gain settings on the deviation from a mono-exponential
signal b-dependence is found. In particular, the deviation occurs at a signal
well above the noise floor. OLS fits of free water provide an unphysical K
value >0 and statistically different results from NLS for anisotropic phan-
toms. WLLS results of MD, MK and FA are equivalent of those of NLS but
with a great spare of time. In biological phantoms, MD is 20% higher with
the K term. MD and FA from the kurtosis fit are more robust among dif-
ferent acquisition protocols.
Conclusions: We demonstrate that a pseudo kurtosis effect can be ob-
served even in free water. For reliable MD estimation, we propose the use
of simple home-made phantoms for quality control, sequence optimiza-
tion and fitting methods comparison. Our data show that the K termmakes
MD more robust; the b-range has to be properly chosen and the WLLS has
to be preferred.
E.449
THE IMPORTANCE OF PERFORMING AN ACCURATE IMAGES
PREPROCESSING IN MR DIFFUSION ANALYSIS
C. Pinardi C. Ambrosi M. Maddalo F. Dusi A. Duina
L. Mascar R. Gasparotti R. Morett a Medical Physics Unit, Hospital
Spedali Civili, Brescia, Italy; b Neuroimaging Lab, Medical and Surgical
Specialties, Radiological Sciences and Public Health, University of Brescia, Brescia,
Italy; c School of Medical Physics, University of Milan, Milan, Italy;
d Neuroradiology Unit, Hospital Spedali Civili, Brescia, Italy; e Medical and
Surgical Specialties, Radiological Sciences and Public Health, University of
Brescia, Brescia, Italy
Introduction: Despite the great success of diffusion MR imaging, the im-
portance of images preprocessing is often underestimated and the review
of data analysis’s pipeline could be overlooked. In order to prove the dif-
ference in the final results between a naive analysis and an accurate one,
diffusion data were processed in two different ways: an easy pipeline and
a state of the art pipeline, according to the discussion held at the diffu-
sion group at the ISMRM 2015.
Materials and Methods: Ten healthy subjects (6 males, 55.2 ± 7.0 y) were
acquired on 1.5 T Siemens Avanto MR unit with a dedicated 8-channel head
coil. The first pipeline used for data analysis is mainly composed of the gui
FSL 5.0.8 tools and represents an easy approach to the preprocessing: a re-
alignment of the two acquired series made with Flirt, a mean of the two
corrected images and a correction for eddy currents made with EddyCorrect,
which performs a linear correction. The second pipeline is more accurate
and uses only the newest algorithms not yet included in the gui FSL 5.0.8
tools: a motion correction of the two series with MCFlirt (specific inline
tool for the motion correction), a mean of the two series and a final cor-
rection for eddy currents made by Eddy (a tool which performs a nonlinear
correction), followed by a rotation of gradient vectors made with Matlab.
e132 Abstracts/Physica Medica 32 (2016) e124–e134