Implementing diffusion-weighted MRI for body imaging in prospective multicentre trials: current considerations and future perspectives

Over the last decade significant hardware improvements have enhanced data acquisition. Signal-to-noise ratio [SNR] improvements have resulted from higher field strength (3T), improved magnetic field gradient performance (increased maximum gradient amplitudes and ramp rates), improved digital radiofrequency (RF) chains and receiver technology with multiple receiver arrays. Advanced digital compensation schemes further mitigate gradient-induced eddy currents reducing image distortion and blur. Although DW-MRI at 3T initially struggled to match the quality of large field-of-view (FOV) 1.5T DW- MRI images because of inhomogeneity of the static magnetic field (B0), recent advances in automated correction (shimming) and improvements in static field homogeneity have made modern 3T platforms viable options for body imaging. In normal volunteers, ADC values of upper abdominal organs are comparable across field strengths; however, the coefficient of variation, (CoV) of the liver was 1.5 - 2.0 times greater at 3.0T compared to 1.5-T [11], emphasising that suitability for inclusion in a multicentre trial requires assessment of individual scanner performance.

Protocol optimisation is often scanner-specific as available measurement and artefact [12] reduction techniques vary between manufacturers, models and software versions. Geometric distortion associated with the static magnetic field can be reduced by using methods to correct field inhomogeneity (advanced shimming) and by increasing the readout bandwidth [13‐15]. Distortions arising from eddy-currents can be diminished by reducing the diffusion-weighting (maximum b-value) and other sequence parameters (echo-train length, matrix) or by employing gradient schemes such as the twice-refocused spin echo [16] that compensate for eddy-currents, as well as by using post-acquisition image registration routines [17]. Ghosting artefacts (displaced re-duplications of the image) can be reduced by adjusting the receiver bandwidth and echo time.

Depending on the disease, an optimal selection of b-values [18] is needed with considerably more b-values required if the signal decay with increasing b-value is to be fitted to non-mono-exponential functions [19]. To avoid confounds from perfusion, b-values of <100 s/mm² should be avoided: maximum b-values of 800 to 1000 s/mm² are usual in body applications (Fig. 1) [20] but their range may need optimisation for specific tumour types. Noise characteristics influence the maximum b-value used in practice. The number of signal averages may be increased at higher b-values to increase SNR [21]. Most DW-MR images are acquired in free-breathing, averaging the signal over physiological motion. Respiratory triggering, using bellows or a navigator, has not shown advantages over multiple averaged free-breathing in estimation of ADCs in abdominal organs [22, 23]. Cardiac triggering has been explored in the upper abdomen [24]. Anti-peristaltic agents reduce image blur arising from peristaltic motion in abdominal and pelvic DW-MRI and multishot techniques may offer some advantages over single-shot techniques in reducing distortion from air within bowel [25].

Parallel imaging reduces geometric distortion, but reduces SNR. The extent of the imaging volume along the scanner bore (z-axis) should be limited to around 25 cm (depending on scanner capability) to mitigate bias in ADC estimates due to spatial non-linearities in diffusion-encoding gradients [26]. For larger volumes, multiple imaging stations can be acquired at the isocentre of the magnet sequentially [27]. Acquisition of multiple stations requires software tools to normalize station-to-station signal variation and the ability to compose the images into a single series for a given diffusion-weighting (b-value). At 1.5T, spectral fat-suppression techniques are often used for abdominal, pelvic or small FOV applications, while inversion recovery is used for whole-body DW-MRI and in regions of poor static magnetic field homogeneity. Fat suppression at 3T is more challenging, and the preferred method may vary between scanners; combinations of suppression techniques may be required [28]. Some consortia such as the Quantitative Biomarkers Imaging Alliance (QIBA) and the European initiative Quantitative Imaging in Cancer-Connecting Cellular Processes to Therapy (QuIC-ConCePT) have been working on standardisation and optimisation of DW-MRI acquisition protocols, and technically validated protocols, e.g., in liver and lung are available to the public [29, 30].

In multi-centre trials, compromises may be required in acquisition parameters in order to achieve an acceptable degree of standardization whilst maintaining good image quality on all scanners [12]. A current list of multicentre trials reporting DW-MRI as a readout in body imaging applications is listed in Table 1.

According to metrology standards of quantitative imaging biomarkers (QIB) [39], measurement performance should be evaluated by assessing repeatability, reproducibility, linearity and metrics of bias. Test-object measurements yield practical estimates of the bias and the repeatability of each clinical MRI system and can be used to compare technical accuracy across the systems [40]. Precise measurement of ADC is important since the dynamic range of the biomarker is quite small, from approximately 0.5×10^-3 mm²/s in densely packed cells to 3×10^-3 mm²/s in fluid-filled cysts.

Ice-water test-objects comprising multiple tubes with distilled water at 0°C and one of sucrose solution [30, 41] have been used but did not provide a sufficient range of ADC estimates. Following this, an ice-water test-object containing multiple sucrose samples doped with metals to reduce relaxation times to physiological values was presented [42] and utilized [12] for optimising a diffusion-weighted protocol in a multi-centre setting. Solutions of polyvinylpyrrolidone (PVP) in water embedded in an ice-water filled sphere [43] or cylindrical vessel [44], remain limited in their range of ADCs (Fig. 2). More specific test-objects have assessed ADC uniformity [12], ghosting and distortions [45, 46].

Test-objects at room temperature are more convenient to prepare than those with ice-water and have been used in single-centre studies [47, 48] but require correction from a temperature-controlled experiment [47] to account for temperature dependence of ADC. The performance of routine test-object evaluations in multi-centre trials involving DW-MRI, their frequency and pass-fail criteria, depends on the trial design and the nature of the imaging endpoint. Test-objects with the required range of ADCs need to be supplied and utilized at participating centres.

Test-objects lack the necessary variation in tissue structure, geometry and motion experienced when imaging humans. Therefore, several trials have built in normal volunteer assessments during set-up.

Optimisation [49] and refinement of the DW-MRI measurement, e.g., multiband techniques [50], Zonal Oblique Multislice [51] and diffusion tensor imaging (DTI) can be assessed [52]. Protocols can also be tailored to underlying tissue structure e.g., lower b-values in pancreas [53] and tolerable versions for clinical use can be developed. Data analysis methods can also be optimised by seeking the best model fits of data from normal tissue [54] which can then be used as a comparator with pathological tissue to investigate structural differences [55].

Healthy volunteer studies are also useful for establishing physiological variation and reference values for disease e.g., in liver [56], bladder [57], bone marrow [58] and breast [59, 60].

Finally, normal volunteer studies are invaluable for studying technique repeatability: coefficient of variation of mean or median ADC estimates in breast 8%,[61] in liver 5.1% [62] and in skeleton 3.8% [63] have been reported. Inter-scanner reproducibility of volunteer data in neurological [64] and abdominal [11] applications provides re-assurance that, with standardisation DW-MRI is suitable for use in multi-centre clinical trials.

A contemporary data archiving framework (termed a Research PACS [65]) needs to consider three important areas:

The extensible Neuroimaging Archive Toolkit (XNAT), an open-source platform (Neuroinformatics Research Group, Washington University, St. Louis,MO,USA) has recently gained significant traction among academic groups as the foundation for such a Research PACS. However, several dedicated clinical trial management systems are also available commercially. Whichever product is used, standard operating procedures (SOPs) must be developed and used for staff training with both trial protocols and legislation vis-à-vis data handling.

Figure 3 presents a schematic of the workflow adopted within multicentre imaging trials. Clear organisation of multiple data types in a central hub brings significant time-savings when retrospective analysis is required [68] and all-electronic data transfer is now rapidly superseding the former practice of posting digital video discs (DVDs) containing trial images. Information governance is implemented via the use of designated staff who exercise a “gatekeeping” role. Data anonymisation by removal and/or replacement of metadata fields in the DICOM files requires a technical understanding of the processing to be done as well as knowledge of trial design and legal expertise. Data protection is achieved by designing robust systems, often including an element of geo-spreading, whilst prevention of unauthorised access is achieved by restriction on an IP address (implemented via appropriate firewall rules), user authentication and role definitions within database software. If a patient withdraws consent, it is possible to remove completely the data from the cohort used for ongoing analysis, but it is likely to prove impossible to remove these data from any summary statistics that have already been published, or any data record deposited as part of the publication process. Government bodies have guidelines pertaining to procedures required to ensure data integrity and compliance with information governance legislation [69].

As the variability of the measurement at low diffusion-weightings is high [70] and the signal decay is exponential, a low b-value of 100-150 s/mm² is preferred when fitting a monoexponential function to derive ADC to reduce the influence of perfusion or flow effects on the measurement (Fig. 1B). Computed DW-MRI, (e.g., b=2000 s/mm²), improves DW-MRI contrast without any measurement penalty [71] but does not contribute to quantitation.

In DW-MRI, the use of non-mono-exponential models (stretched exponential, kurtosis, statistical and bi-exponential) [72‐76] probe aspects of tissue microstructure [77] and differences between tumour sub-types or inter-tumour heterogeneity [78‐83]. They may also provide an earlier indication of response to treatment than ADC estimates [84, 85]. Selection of the most appropriate model remains an area of active research: use of a model with many additional parameters risks over-fitting the data and may be sensitive to noise characteristics of the system rather than structural properties of the tumour or normal tissue. Vendor-supplied software to support calculation of these alternative diffusion attenuation models would help address some of these issues [77‐86].

Finally, retention of tumour segmentations allows quality control (QC) review of data reduction procedures, as well as facilitating retrospective trial of alternative diffusion metrics drawn from the same 3-D segmentation objects stored at the pixel level [87]. As interobserver concordance is dependent on extent of sampling [88], the method of segmentation should be clearly recorded, for example, whether whole tumour or selected slices are segmented, and whether necrotic or cystic areas are excluded. A manual, semi-automated or automated method could also introduce variability in the measurement [89] and should be standardised.

Following set-up and Quality Assurance (QA), tests should be carried out at the beginning of the study to assess the baseline performance of each scanner, followed by regular QC tests over the course of the study (particularly after servicing and software upgrades) to detect changes in performance (Table 2). The frequency of tests and defined action limits, which specify the range of acceptable values may be study-dependent.

Within a multicentre trial, QA and QC procedures for imaging depend on the role of imaging in the trial [90]. Qualitative interpretation does not require the same level of QA/QC as for deriving quantitative data. The ROI size and number of pixels within it are crucial for quantitative assessments, particularly as many studies now address ADC distribution rather than mean or median values. Operational support for imaging QA and QC should be in place at trial setup and through the life of the trial (Table 2). A standardised and optimised acquisition protocol, which acknowledges vendor differences and incorporates acceptable and non-acceptable deviations should be defined and supplied to sites upfront. Acquisition of test data (test-objects, volunteers) reduces the likelihood of poor quality or non-evaluable imaging data being acquired from the first patient in the study; occasionally the first 1 or 2 patients may be considered as “run-in” to assess site compliance and data quality. From an ethical perspective, the intention must be for all included patients to contribute analysable data. However, if sites find it difficult to comply with the protocol, or if the first few patients' data are of poor quality, it may be necessary to discard those data following a protocol amendment to improve the methodology. Prospective QC with timely and informative feedback to the site enables supplementary correction to be taken and avoids non-assessable poor quality data at the end of the trial. Site upload of anonymised data via a web-based system requires training so that data are securely handled and correctly coded for inclusion in the trial imaging database.

Measurement uncertainty arises from differences in acquisition (hardware and software differences between scanners as well as within scanners variations due to use of different protocols) plus post-processing parameters, longitudinal changes or ‘drift’ in MRI signal when using the same scanner over the study period as well as from natural physiological variation within and between study participants. The Radiological Society of North America (RSNA) Quantitative Imaging Biomarkers Alliance (QIBA) recommends that evaluation of biomarker reliability includes analysis of precision and bias estimation, plus measurement linearity, by comparison with an accepted reference or standard measurement [91]. For DW-MRI, in vivo physiological references are not available for bias/linearity measurements and these are extrapolated from phantom studies [20, 39, 91, 92].

Assessment of technical performance of an imaging biomarker includes measurement variability arising through differences between scanners (same patient, different scanners) [11], imaging protocols [93] and post-processing methods (such as different analysis software, lesion segmentation methodologies [94] and imaging readers [91, 92, 95]).

In trial design, the context in which the biomarker is being utilized dictates the measurement variations that must be accounted for. If measuring therapy-induced change, where it is usually possible to image each patient on the same scanner and for all analysis to be carried out by the same investigator, precision estimation is limited to repeatability [39]. For studies aimed at prognostication or lesion characterisation, ADC values will be compared between individuals or across institutions and as it is necessary to know whether a measured difference represents a true difference, measurement uncertainty including statistical appraisal due to reproducibility must be evaluated.

Coefficients of variation at different anatomic locations are in the range 3-10% [20, 96]. Inter-vendor two-site reproducibility coefficients of variation range from 14-27% [20]. In multicentre trials, a measured difference should be outside the 95% limits of agreement of the measurement uncertainty expected in a multicentre trial setting for it to be attributed to a true treatment-related difference. Alterations in lesion geometry also may affect segmentation thresholds and need consideration when making longitudinal measurements [97].

Clinical trials of investigational drugs and devices must comply with International Conference on Harmonisation GCP if they are intended to support regulatory approval [98]. For multicentre imaging studies, challenges exist in ensuring that different makes and models of MR scanner yield comparable data [90] and maintaining compliance with unfamiliar protocols at trial centres. The Food and Drug Administration has specific guidelines to help ensure that imaging biomarkers are measured in accordance with the trial’s protocol [99], and that quality is maintained over time and between sites: it recommends that sponsors employ an “Imaging Charter”, ancillary to the trial protocol, which defines the imaging process in exhaustive detail. Sponsors often engage specialist Imaging Clinical Research Organisations to perform site qualification and training, phantom-based QA/QC, pilot studies, data management and analysis. Double baseline studies are valuable in verifying repeatability [100], although the additional burden may deter patients, sponsors and ethical committees.

Performing imaging in clinical trials risks discovery of incidental findings (IFs) that may require action and, therefore, require review by a trained diagnostician [101]. Ethical and legal issues surrounding IFs are a key element of the duty of care owed by researchers to study participants (Table 3). Generic recommendations are offered by the National Institute of Health in the USA and Royal Colleges in the UK (Table 3). No specific recommendations have yet been proposed for studies utilising DW-MRI.

A report of whole body DW-MRI in healthy volunteers has shown IFs in 29% of subjects. Of these 30.6% were considered of ‘moderate significance’ and 10.2% ‘high significance’, requiring specialist review but only a minority of scans required further action [102]. In myeloma, IFs were seen in 38% (67/175) of examinations, 20% of findings were equivocal and after specialist radiologist and clinical review, only 3% of cases prompted further investigation. It is mandatory to introduce an image review process, triage and referral pathways embedded into trial design and reflected in consenting procedures. For multicentre trials, this system should account for the logistical hurdles that arise due to data storage and delays in data viewing. For cases where data are interpreted centrally, procedures should define a reporting mechanism, so that IFs discovered centrally prompt action locally.