INTRODUCTION
In cancer studies of several organs (e.g. breast, prostate, liver, etc…) functional imaging can give useful information for better quantify the staging. In particular, Diffusion Weighted Imaging (DWI) and Dynamic Contrast Enhanced (DCE) MRI are becoming increasingly adopted in clinical practice.
DWI can give information about the microstructure organisation of the tissue. In fact, tumors have an increased cellularity leading to reduced water diffusion.1 Isotropic diffusion can be modeled using Intra-Voxel Incoherent Motion (IVIM)2 which depends on a few parameters: perfusion fraction f (fraction of plasma vessels), ‘slow’ diffusion Ds (conventional diffusion), ‘fast’ diffusion Df (diffusion in randomly oriented vessel). Diffusion constants are measured in mm2/s.
DCE provides information on the capillary wall permeability. This latter increases in cancer because of leaky vessels associated to the fast grow induced by VEGF (vascular endothelial growing factor).3 One of the most commonly employed models is the Extended Tofts-Kety (ETK) model4 includes the following parameters: vascular plasma fraction (vp), transfer constant from plasma to extra-vascular-extra-cellular space (EES) Ktrans, transfer constant from EES to plasma kep. Transfer constant are mesured in min-1.
Both models parameters are typically estimated from measured DCE and DWI data. Due to noise the estimated parameters can inaccurate and imprecise.
Following the suggestion by LeBihan et al.5 in this study we hypothesized that perfusion fraction (f) in IVIM and plasma volume fraction (vp) should be equal and might be used for improving the estimation of parameters attainable with DWI and DCE.
In particular, we performed a Monte Carlo simulation in order to evaluate how the IVIM and ETK fitting can be performed simultaneously and if the simultaneous fitting can produce better performance than separate fitting.
METHODS
Intra-Voxel Incoherent Motion (Ivim) Modeling
The IVIM model has been proposed by Le Bihan et al2 according to this model the signal intensity of the diffusion weighted images is given by: where S0 is the signal intensity with no diffusion gradient, f is the perfusion fraction, Ds is the slow diffusion coefficient of water and Df is the ‘fast’ diffusion coefficient, b is the b-value that is related to the pulse magnetic gradient intensity and duration.
Extended Tofts And Kermode (Etk) Modeling
Tofts and Kermode proposed a model for the capillary flow.4 According to this model the concentration of contrast agent within a voxel is given by: where t is time, Ktrans min-1. is the transfer constant across the capillary wall from the plasma to the Ees, kep is the reverse constant from EES to plasma, vp is the plasma volume fraction.
Simultaneous Fitting Of Ivim And Etk
Le Bihan et al5 suggested that the perfusion fraction within the IVIm model should be interpreted as the vessel fraction within a voxel. This hypothesis, might be used to improve the numerical estimation of model parameters (DWI and DCE) performing simultaneous fitting of both data.
In order to take advantage of the link between IVIM and ETK models we need to perform a simultaneous fit of the data. In this study we propose to perform minimization of an appropriate cost function. Under the hypothesis that f=vp the cost function can be written:
where is the signal acquired during DWI and is the signal acquired during DCE-MRI.
SIMULATIONS
We simulated one ETK curve with Ktrans=0.046 min-1. and kep=0.148 min-1. and vp=0.0074. The parker model for the arterial input function has been used.6 Time sampling was simulated as in fast TWIST (Time-resolved angiography With Interleaved Stochastic Trajectories) pulse sequences with an image every 3 sec. Noise was superimposed having a standard deviation of 0.1 mMol.
Further we simulated a IVIM curve using f=vp=0.0074, Ds=0.001 mm2/s, Df=0.01 mm2/s and b-values (0, 50, 100, 150, 200, 500, 600, 700, 800, 900, 1000, s/mm2). Noise was simulated having standard deviation 0.05 normalised units.
Simultaneous fitting using the cost function proposed in the previous section has been performed. Starting values were: Ktrans=1, kep=0.5, vp=f=0.5, Ds=1e-3, Df=10e-3. The Levenberg- Marquardt algorithm in Matlab has been used.
Monte Carlo simulation on 10 repetitions has been performed on the same noisy data but with different starting points chosen randomly within 10% of the starting values reported in the previous paragraph.
RESULTS
Figures 1 and 2 report the simulated curves with superimposed noise. Figure 3 report the result of fitting using the proposed cost-function (Figures a and b), and separately fitting the data (Figures c and d).
When using simultaneous cost-function, mean relative error (expressed in percentage %) on parameters were: Ktrans=-7.20%, kep=11.54 %, vp=f=21.01%, Ds=-5.1132%, Df= -99.9850%. When the fitting were performed separately mean percentage errors were: Ktrans = -3.75 %, kep=6.37% vp=4.25% for the ETK model and f=6*106%, Ds=6.13%, Df= 347%.
DISCUSSION
In this study we explored the possibility to perform simultaneous fitting of DCE and DWI data under the hypothesis that plasma fraction and perfusion fraction are the same as suggested by Le Bihan et al.5 We performed Monte Carlo simulations with noisy DCE and DWI data. Simulated data were fitted separately and in combination.
Results suggest that DCE fitting can lead to better results when performed alone. However, DWI data fitting can lead to huge errors in parameter estimates that might be attenuated when fitted simultaneously with DCE data.
Although results seems promising, we must underline the limits of this study. In our study we did not evaluate the impact of different curve fitting algorithms, the influence of the noise level, the influence of the number of b-values used.
It seems reasonable that a larger number of b-values might improve DWI fitting alone and consequently also the combined fitting.
CONFLICTS OF INTEREST
None.