This poster is published under an open license
. Please read the disclaimer
for further details.
Computer applications, Oncology, MR-Diffusion/Perfusion, Computer Applications-Detection, diagnosis, Neoplasia, Abscess, Metastases
G. C. Manikis1, K. Nikiforaki1, G. Ioannidis1, N. Papanikolaou2, K. Marias1; 1Heraklion/GR, 2Lisbon/PT
Methods and materials
Synthetic bi-exponential data of predefined organ related set of initial values for D,
and D* were used to be approximated by the IVIM equation and the NLLS. The ground truth values reflect a wide range of D,
and f as presented in the literature for several organs in order to better approximate the real tissue characteristics [2-4].
Six different Rician noise levels were added to the synthetic data,
producing diffusion signals of equivalent quality of the ones produced in clinical routine.
The estimated lowest to highest SNR values were grouped into six distinct intervals of equal population (SNR1-SNR6).
To assess repeatability,
each noise level was randomly added ten times to each of the signals. Finally,
a total of 47.520 synthetic diffusion signals contributed to the analysis.
To further demonstrate the performance of the proposed fitting method in clinical practice,
this work concludes by using clinical datasets of the same imaging parameters used to assess abdominal malignancies in a 1.5T scanner.
Standard Fitting Scheme
The current work was based on the non-linear least squares (NLLS),
one of the most widely used minimization method for fitting the IVIM model.
as an iterative procedure,
needs a starting point and the parameter’s natural lower and upper bound that limit its calculated value and reach the optimum fitted value.
Standard fitting techniques using NLLS includes: a) using NLLS for direct estimation of D,
and f from the IVIM model (complete standard fitting-CS),
and b) calculate D linearly at high b-values (i.e.
b>200 s/mm2) and the remaining parameters f and D* by the NLLS and the IVIM model (partial standard ftting-PS).
Proposed Fitting Scheme
The proposed optimization scheme (complete proposed fitting-CP and partial proposed fitting-PP respectively),
acts as a plugin tool before CS and PS,
aiming to define initial,
lower and an upper bound for each IVIM parameter.
Initial estimation of D was achieved by fitting linearly the diffusion data in a logarithmic scale.
With known initial value for D,
initial value of f was then calculated using the exponential part of the IVIM model reflecting the true-diffusion part.
the initial value of D* was obtained using the IVIM bi-exponential model.
Lower and upper bounds of the IVIM parameters were automatically assigned with respect to their initial values.
Adjusted R-square (adj-R2) was chosen as the goodness-of-fit metrics to assess fitting accuracy from the four fitting schemes.
Relative errors between the ground truth and the fitted IVIM parameters were also calculated. Repeatability test was implemented under standard widely used metrics including the coefficient of variation (CV),
and the intraclass correlation coefficient using two-way mixed single measures (ICC(3,1)).
Both CV and ICC(3,1) were calculated using the ten synthetic data for each predefined IVIM group of parameters and the measured SNR value over the four fitting methods.