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

139 / 146

tering algorithm). Dice and XOR Overlap coefficients were estimated first
for all 37 patients (case 1) and then without considering 15 healthy pa-
tients, confirmed by EAM and with artefacts in the DE-MRI images (case
2). Differences of coefficients between the two cases were tested with
Wilcoxon–Mann–Whitney test.
Results: From case 1 to case 2 XOR Overlap coefficient has no significant
variation of median value and interquartile range (m1 = 0.94 to m2 = 0.96,
IQR1 = 0.08 to IQR2 = 0.05, P = 0.09) whereas for Dice coefficient, due to in-
correct classification of fully automatic algorithm used in case 1, there is
an increase in median value and decrease in the data dispersion (m1 = 0.44
to m2 = 0.73, IQR1 = 0.46 to IQR2 = 0.31, P < 0.001).
Conclusions: This study shows the relevance of robustness of 3D LA fi-
brosis segmentation based on DE-MRI and its reliability in clinical practice.
The fully automatic k-means based algorithm is not as robust to the pres-
ence of DE-MRI images artefacts as the semiautomatic threshold algorithm.
E.453
ACCURACY COMPARISON OF TWO DIFFERENT OPTIMIZATION
ALGORITHMS FOR DIRECT BI-EXPONENTIAL ANALYSIS OF IVIM MRI
IMAGES
P. Volonghi L. Mascar L. Spiazzi G. Rizzo R. Moretti a Institute of
Molecular Bioimaging and Physiology (IBFM), CNR, Segrate (MI), Italy; b U.O.
Fisica Sanitaria Spedali Civili di Brescia, Brescia, Italy
Introduction: Intravoxel incoherent motion (IVIM) MRI technique allows
the estimation of diffusion coefficient (D), perfusion fraction (f) and pseudo-
diffusion coefficient (D*), related to perfusion and diffusion of tissue water
molecules. The estimation of these parameters is usually obtained through
a bi-exponential analysis of image signal as a function of b-values. The aim
of this work was to compare the parameters estimation accuracy using two
different optimization algorithms for direct bi-exponential fitting.
Materials and Methods:Parametric images, created using parameters values
referring to literature, were altered with noise derived from a temporal series
of acquisitions of a homogeneous phantom, which was scanned with a
Siemens Aera 1.5T using EPI sequences with increasing b-values (0, 50, 100,
150, 200, 400, 800, 1000 s/mm2). Considering a ROI inside the phantom,
noise images for each b-value were obtained for each pixel by filtering of
the temporal trend dataset. The IVIM parameters were estimated using a
direct bi-exponential fitting of the decaying signal, testing two optimiza-
tion algorithms (available in Matlab): Levenberg–Marquardt (LM, the most
used in literature) and Trust Region (TR). The algorithm accuracy was evalu-
ated for each parameter in terms of error percentage.
Results: Regarding D estimation, the error mean was lower and less vari-
able for TR (4.1 ± 3.1%) compared to LM (5.2 ± 5.3%); in case of f estimation,
TR and LM accuracy was similar (TR = 16.6 ± 17.2%, LM = 19.3 ± 35.5%), while
for D* estimation, TR accuracy was strongly better than LM accuracy
(TR = 41.0 ± 60.4%, LM = 224.7 ± 3467.7%).
Conclusions: Direct bi-exponential fitting of D parameter had generally a
good accuracy, which was slightly better using TR algorithm. Regarding the
estimation of f and D* parameters, the accuracy decreases, in particular,
the two optimization algorithms gave comparable results for perfusion frac-
tion parameter, while TR had better performance than LM for pseudo-
diffusion coefficient.
e134 Abstracts/Physica Medica 32 (2016) e124–e134