Intravoxel incoherent motion (IVIM) MR imaging has been emerging as a potentially plausible technique that can quantitatively assess both water molecular diffusion and capillary blood perfusion noninvasively in human liver ¹. To date, however, the numerical values reported for the pseudo-diffusion coefficient that reflects liver perfusion, D*, vary to a great extent from 27 to 190mm²/s ^2-4. It is hypothesized that modeling of the signal decaying behavior plays a role in D* quantification, in addition to effects of respiratory motion ⁵. In this study, therefore, we aim to increase the reproducibility of D* measurements via proper selection of the signal decaying model and the b-values. Results on 10 healthy subjects scanned twice showed substantial improvements in inter-scan reproducibility in D*.Derivation of liver D* using IVIM imaging is based on fitting of the signals as a function of b-values using the bi-exponential decaying model: S_b = S₀ [ f*e^-bD* + (1-f)*e^-bD ], where S_b and S₀ are the signal intensities at a specific b value and b=0, respectively, D the true diffusion coefficient, and f the fraction of blood flow in the region of interest (ROI). At least two problems could cause fitting uncertainty. First, in the absence of prominent vasculature, single exponential model suffices, and over-fitting may result when using bi-exponential model. Second, the presence of large vessels may dominate the data at b-values less than 15s/mm², leading to D* over-estimation ⁶. To solve these issues, we first restricted curve fitting using data with b>15s/mm², followed by the combined use of single- and bi-exponential models for fitting, for which the one exhibiting smaller residual fitting error was chosen for the region of interest.All experiments, IRB approved, were carried out on a Siemens 3T scanner with a torso array for signal receiving. Image data at sixteen b-values (0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 150, 200, 400, 600 and 800 s/mm²) were obtained from ten healthy young subjects using respiratory-triggered spin-echo EPI with diffusion sensitizing gradients. Each subject was scanned twice on the same day with re-positioning to examine inter-scan reproducibility. Scanning parameters were: FOV=310x310 or 410x410mm, TE/TR=75/2000ms, thickness=7mm, matrix size=112x112, NEX=6. For estimations of IVIM parameters, circular ROIs at about 1cm diameter were placed on the right hepatic lobe (Fig.1). Curve fitting to derive D* was performed using the proposed scheme described above, with the consistency between two measurements of the identical subjects analyzed using the Bland-Altman plot, and compared with conventional fitting results using solely the bi-exponential model. All data processing was carried out with self-designed Matlab scripts.Figures 2 and 3 show the Bland-Altman plots for D and D*, respectively. The proposed scheme yielded similar level of reproducibility around 19% imprecision (Fig.2a) to that using the conventional bi-exponential model for the true diffusion coefficient D (Fig.2b) and for the perfusion fraction f (not shown). On the other hand, D* reproducibility was improved substantially from 104% imprecision using conventional bi-exponential model (Fig.3a) to 39% imprecision using the proposed scheme (Fig.3b). With our method, D* for the ten subjects mostly fall within the range of 38 to 85x10^-3 mm²/s.Prominent inconsistency in the numerical values for liver D* has been found in past reports employing IVIM imaging ^{2-4, 6}. Results from our study suggest that such phenomena may have originated, at least in part, from inappropriate use of the signal model for curve fitting. Our proposed scheme yielded substantially improved inter-scan D* reproducibility on repeated scans, while precisions in D and f were unaltered. The method may find potential in clinical evaluation of liver perfusion using IVIM MR imaging.