Currently, there are many models available to describe diffusion weighted MRI (DWI) data signal intensity as a function of diffusion weighting (b-values). Among these models are the: classical mono-exponential model (mono-exp), bi-exponential intra-voxel incoherent motion (IVIM)-model, tri-exponential model (tri-exp), stretched-exponent model (stretched-exp), Gaussian-model (Gaussian spread of D-values) and kurtosis-model. Often, the mono-exp and IVIM-models are used, as they are most easy to interpret. However, other models may provide parameters that are either more precise or more sensitive to detect tissue abnormalities.

Therefore, in this research, we evaluated how the abovementioned models perform in pancreatic cancer patients considering the repeatability of the model parameters as well as their potential to differentiate between healthy and tumorous tissue.

We acquired 2D multi-slice diffusion weighted (DW) EPI images in nine patients with either histologically or cytologically proven pancreatic cancer (three females, mean age 69, range 56-77) on a 3T Philips (Ingenia) scanner, using navigator based respiratory triggering. Images were acquired at b=0,10,20,30,40,50,75,100,150,250,400 and 600s/mm². All patients were scanned three times during two separate sessions (1-8 days apart) to assess intra-session and inter-session repeatability. To minimize bowel motion, patients were administered hyoscine bromide (Buscopan; 20 mg iv) right before the first DWI acquisition each session.

All DW-images were denoised using a Rician adaptive non-local means filter¹ and registered group wise² using a 4D non-rigid b-spline algorithm based on mutual information, in Elastix³. All six models mentioned above were fitted to the data, using the Levenberg-Marquardt least squares fitting algorithm in Matlab. From this, maps were created for the various model parameters: (apparent) diffusion coefficient (ADC; D), perfusion coefficient(s) (f,f₁ and f₂), pseudo-diffusion coefficient (D*), exponential stretch (α), standard deviation of the Gaussian-model (σ) and kurtosis (K). For the tri-exp model, the pseudo-diffusion coefficients D*¹ and D*² were fixed to 0.014 and 0.093 mm²s^-1 (based on healthy pancreatic data in 16 volunteers, data not shown).

For the bi-exponential IVIM-model, we used three fitting algorithms: Levenberg-Marquardt least squares (IVIM-Free, as used for all other fits), Levenberg-Marquardt least squares while fixing D* to 50×10^-3 mm²s^-1 (IVIM-Fixed) and the adaptive threshold algorithm⁴ (IVIM-Adaptive).

Per patient, we created two single-slice regions of interest (ROIs), containing either healthy pancreatic tissue or pancreatic tumor tissue. The ROIs were drawn on an ADC-map generated from b=100 and 600 s/mm² under the guidance of gadolinium contrast-enhanced Dixon-reconstructed images. Per patient, we calculated the mean value of the voxel-wise fits within the ROIs for all fit parameters.

Per fitting parameter, we used a paired t-test to test whether the parameter values from the tumorous ROI were significantly different from the values in healthy tissue (p<0.05). In addition, we tested the repeatability of each fit parameter by calculating the inter- and intra-session within-subject coefficient of variation (CV) from the tumor ROI.

For the mono-exp, IVIM, tri-exp and stretched-exp fits the ADC/D-maps looked similar to each other (Figure 1). For the multi-parametric fit models, the non-D parameter-maps showed complementary information with respect to the D-maps, as their hypo-intense and hyper-intense regions were distributed differently throughout the tumor (e.g. f, Figure 1).

Each model had one fit parameter that was significantly different between tumorous and healthy tissue (Table 1, Figure 2). For the IVIM-model fits, the perfusion fraction was the only parameter that differed significantly.

In general, parameters that showed high contrast between healthy and tumorous tissue had a lower reproducibility as indicated by larger CVs (Table 2, Figure 3).

We cannot indicate a preference for a specific model based on both repeatability and sensitivity, as the model-parameter with the highest repeatability had poorer sensitivity compared to other model-parameters and vice versa. When high sensitivity is desired, one should use the f₂ from the tri-exp model whereas when high repeatable is desired, the ADC from the mono-exp model should be used.

Parameters from multi-parametric models showed complimentary information. Therefore, multi-parametric models may be preferred above the mono-exponential model. However, this characteristic was not yet fully exploited in the current analysis.

From the tested IVIM fit methods, there was no clear preference considering repeatability and difference in healthy and tumorous tissue.

Each model had one fit parameter that differed significantly between healthy tissue and tumor, enabling the use of all models as a diagnostic tool for pancreatic cancer. From these parameters, the most sensitive parameter was f₂ from the tri-exponential model, whereas the best repeatable measure was the ADC from the mono-exponential model.