An increase in apparent diffusion coefficient (ADC) measured over the entire disease burden has been shown to be indicative of response to chemotherapy in patients with advanced ovarian cancer.¹ Advanced models may provide a better description of the attenuation of the diffusion-weighted MRI (DW-MRI) signal with increasing diffusion-weighting (b-value)² and may provide an earlier indication of response than ADC.³ Advanced models may, however, be more strongly affected by differences in sequence parameters and noise characteristics between scanners, which may limit their applicability in multi-centre studies and confound comparison between results from different scanners and protocols.

To assess repeatability and inter-scanner reproducibility of fitted parameters from mono-exponential, stretched exponential, kurtosis and bi-exponential models of DW-MRI data in abdominal organs in healthy volunteers and patients with relapsed ovarian cancer imaged on 1.5T MR scanners from three manufacturers.

Study protocol: Healthy female volunteers and patients (details in Figure 1) recruited at three institutions gave written consent to participate in this Institutional Review Board-approved study. Healthy volunteers underwent two DW-MRI examinations at their recruiting institution (median interval between scans: 5 days, range 1 to 8 days). 37 patients with relapsed ovarian cancer with at least one lesion larger than 2cm were recruited as part of an ongoing prospective multi-centre clinical trial (DISCOVAR, NCRN-portfolio number 11182). Patients were scanned once before starting chemotherapy, at their recruiting institution.

Imaging protocol: Hyoscine butylbromide (20mg) i.m. was administered to patients before scanning to reduce image artefacts due to peristalsis. DW-MRI protocols are described in Figure 1. In patients, the imaging volume was positioned on the largest lesion; in healthy volunteers the imaging volume covered kidneys, liver and spleen.

Analysis: Regions of interest (ROIs) were drawn by region-growing on computed DW images (b=1000smm^-2 for lesions; 500smm^-2 for kidneys; 800smm^-2 for liver; 1000smm^-2 for spleen) using in-house software.⁴ In lesions, ROIs drawn on every slice on which the lesion appeared were combined to produce a volume of interest (VOI). In kidneys, liver and spleen, VOIs encompassed the whole area of the organ on three contiguous slices. Mono-exponential$$$\;S(b)=S_0\exp(-b\mathrm{ADC})$$$, stretched exponential$$$\;S(b)=S_0\exp(-(b\mathrm{DDC})^\alpha)$$$, kurtosis$$$\;S(b)=S_0\exp\left(-b\mathrm{D}_k+kb^2\mathrm{D}^{2}_{k}/6\right)\;$$$and bi-exponential models$$$\;S(b)=S_0\left(f\exp(-b\mathrm{D}^*)+(1-f)\exp(-b\mathrm{D})\right)\;$$$were fitted to all 10 b-values for every pixel in the VOIs using least-squares fits (trust-region-reflective algorithm, Matlab 2014a); the kurtosis model was used empirically here and the original theoretical interpretation is not implied.⁵ The median of each fitted parameter from all pixels in the VOI was used for further analysis. Repeatability of median ADC,$$$\;$$$DDC,$$$\;$$$α,$$$\;$$$D_k,$$$\;$$$k,$$$\;$$$D,$$$\;$$$f,$$$\;$$$D* and the compound parameter fD* was assessed using pairs of measurements from kidneys, liver and spleen in healthy volunteers on each scanner. Bland-Altman plots of untransformed data showed a relationship between the differences in repeated measurements and their means that was reduced by using the natural logarithm of the data. The Coefficient of Variation (CV) of the log-transformed data was used to describe the repeatability of the fitted parameters,$$$\;\mathrm{CV}=\sqrt{\exp\left(\Sigma_id_i^2/2N\right)-1}$$$, where$$$\;d_i\;$$$is the difference between paired measurements for volunteer $$$i$$$, and $$$N$$$ is the number of volunteers. For k, the standard deviation of differences between two measurements was used to describe repeatability as k can take negative values. For each fitted parameter, differences between scanners, as well as differences between organs/tumour sites, were assessed using two-way analysis of variance (ANOVA, Matlab 2014a). A threshold of p<0.006 was used (i.e. 0.05/9) to account for multiple comparisons.

The repeatability of ADC was good, with DDC,$$$\;$$$α,$$$\;$$$D_k and D exhibiting comparable repeatability to ADC on all three scanners (Figure 3). The standard deviation of differences in k was comparable across three scanners. The repeatability of f,$$$\;$$$D* and fD* was poor on all scanners. Data from healthy volunteers showed a significant difference between organs in all fitted parameters and a difference between scanners in α,$$$\;$$$k,$$$\;$$$D and f (Figure 4). In lesions, there was a significant difference in k between scanners (Figure 5).

The comparable results obtained in ADC estimates from different scanners indicates that ADC is suitable for use in multi-centre studies. The differences between scanners in parameters from stretched exponential, kurtosis and bi-exponential models indicate that these models may be more sensitive to differences between scanners, although significant differences between organs could still be detected. Although data from patients did not show differences between scanners in any parameters except k, this may be due to greater heterogeneity between lesions, which may mask inter-scanner differences, whereas differences were significant in the tightly-controlled volunteer cohort.

ADC estimates obtained using 1.5T MR scanners from three manufacturers show good repeatability and no significant differences between scanners. Significant differences between fitted parameters from advanced models should be considered when employing DW-MRI in multi-centre studies.