Diffusion data were post processed by an in house software (ICS,
FORTH) which is able to produce pixel based parametric maps of a number of model related parameters.
All pixel values belonging to the tumor,
as marked by a radiologist,
were used as input for signal intensity curves as a function of b-value.
Model specific curves were grafically overlaid on the data in order to gain insight into each model performance qualitatively.
For quantitative evaluation of each approach,
statistical metrics permitted direct comparison of the fitting outcome.
Μetastatic liver can serve as a good candidate to test a complete fitting model with adequate sampling in the low (b min:50) ,as well as in the high b value area (b max:3000),
for an underlying twofold reason: A.
Normal appearing liver parenchyma (Fig.
as a tissue with dense cappillary network,
can challenge model performance in areas of increased microperfusion contamination of the low b value diffusion signal and B.
Metastatic (-necrotic) tissue,
characterized by anomalous structure (Fig.
can elicit deviations from normal gaussian distribution of diffusivity,
prominent in data acquired in the high b (> 1000) value area.
An indicative analysis result,
when the goodness of fit of the four models was assessed by the adjusted R-square,
is outlined in Fig 3.
Boxplots were generated according to the adj-R2 value in every pixel in the examined region,
showing a clear evidence that the bi-exponential models in general better fit the data of each patient.
bi-exponential non-Gaussian model was proved to fit the experimental data better than any other model according to 3 out of 4 statistical measures (adj-R2,
RMSE and SSE,
p<0.01 in all pairwise comparisons).
standard deviation and square error of each statistical measure is presented in Tables 1-4.
The cAIC showed no clear differences except that the non-Gaussian bi-exponential model was less accurate than the other 3 models.
The imaging biomarkers derived from the four models are quantitatively presented in Tables 5-7.
Biexponential non-Gaussian model fitted the experimental data from hepatic metastases and normal liver,
better than any other model according to all 3 statistical measures (adj R square,
RMSE and SSE,
p<0.001 in all pairwise comparisons).
Adjusted R square of nonGaussian biexponential model of hepatic metastases and normal liver were 97.26% and 97.16%,
while adjusted R square of monoexponential Gaussian model of hepatic metastases and normal liver were 90.3% and 88.7%,