Consensus-based technical recommendations for clinical translation of renal BOLD MRI

During the PARENCHIMA Annual Plenary Meeting in Prague (October 4–5th 2018), the Task Force on Technical Recommendations for Clinical Renal MRI decided to follow a modified Delphi method [30‐32] for seeking consensus regarding technical BOLD MRI acquisition and analysis. The Delphi method derives consensus recommendations through qualitative evaluation of published studies by experts. Where published data is scarce, experts can make inferences based on data from similar contexts, or their own experiences. Published evidence is summarized in the form of a questionnaire. Participants are invited to respond to the questionnaire in several rounds (typically 3), reviewing an anonymous summary of the previous responses before responding to the subsequent rounds. Respondents are asked to indicate whether they consider themselves sufficiently knowledgeable to answer each of the questions. Consensus on a question is usually defined as 68%-80% unanimity in responses [31‐33]. Discussion in a face-to-face meeting between all respondents usually follows one or several of the rounds, and serves to build consensus in the areas where prior responses failed to achieve consensus. Owing to wide geographic distribution of the experts and respondents, we have adopted a modified Delphi method for this project with only a single face-to-face meeting before the final round of surveys.

The questionnaires were constructed by the panel co-chairs (PVP, OB and SF), between February 2019 and April 2019, and administered online to participants in two rounds via an online form (Google Forms, https://​docs.​google.​com/​forms). In constructing the questionnaires, the co-chairs used a recently published PARENCHIMA statement paper [29] and its detailed summary of parameters in the supplementary material, to formulate questions on renal BOLD MRI research topics, technical image acquisition parameters and analysis methods. Round 1 consisted of 35 questions with multiple-choice responses, with the possibility to abstain or to provide a long form answer. Participants were prompted to explain the reasoning behind their choices in a comment section following each item, and also to suggest questions for inclusion in Round 2 of the questionnaire.

Six members of the panel attended a face-to-face meeting between March 18th–19th 2019 at Aarhus University Hospital, Denmark. Based on a review of the first round of questionnaire responses, they worked together on additional questions and on the wording for specific questions in Round 2. Round 2 consisted of 31 items with answers in simplified multiple-choice format (“I agree”, “I disagree”, “I have insufficient experience to make a recommendation”) and possibility to explain the choice in a comment section following each item. Questions that reached consensus in Round 1 were summarized to the respondents in Round 2, and respondents were offered opportunity to comment if they did not agree. Respondents were encouraged to answer the questionnaire based on published evidence (some items pointed to specific publications) and best practices as reflected in the literature, not necessarily on their current research practice. Both iterations of the questionnaire had items grouped under four themes: (1) patient preparation (common items among all PARENCHIMA task force panels for Round 2, see covering review paper); (2) acquisition; (3) analysis; (4) application and interpretation.

The first panel consisted of 10 PARENCHIMA task force members with a publication record and research interest in BOLD MRI. To expand our pool of respondents for the 2nd round, we invited other experts outside of our collaboration. We identified renal BOLD experts on the basis of a PUBMED literature search using [BOLD MRI] AND [kidney] AND [human]. An additional 16 respondents were enrolled while two respondents to the first questionnaire did not participate in the second one. For the second and final round of the questionnaire, we had 24 respondents from 22 academic institutions, 9 countries and 3 continents. All respondents are listed as authors of this manuscript. Panelists were asked to list their specialty when responding to the questionnaire.

At the Aarhus meeting, a “traffic light” system was adopted to issue recommendations based on the degree of consensus achieved through the questionnaire. Green light (consensus, closed issues) on a topic was defined as at least 75% unanimity in responses to a question. An “orange light” was defined in the case where the responses showed clear opinions (65–74%), but did not meet the threshold. For a “red light” topic, the questionnaire showed no clear expert preference (50–64%), so that no recommendation is possible with the information given to the respondents. It was also agreed that when calculating percentage of responses, the “abstain” responses (i.e. “I have insufficient experience to make a recommendation”) will be excluded. However, the percentage of “abstain” responses for each item will be reported, to reflect the level of familiarity of the experts with the topic and to show the topic’s importance (or lack thereof) in the clinical renal BOLD MRI community.

The final results of the survey are summarized for each area in Tables 1, 2, 3, 4, 5, 6 and 7. For the second and final round of the survey, physicists (MR physicists or biomedical engineers) and clinicians (radiologists, nephrologists or physiologists) were equally represented (Table 2). The panelists reached consensus on 14 out of 31 items (45%) (Tables 3, 4, 5, 6). Additionally, panelists concurred on the nine questions for which consensus was achieved in Round 1 (Table 7). 12 items met the “orange light” definition, showing a clear direction of respondents’ opinions, while 5 items were “red light” topics with a 50/50% split in respondents’ opinions. Even though the overall response did not reach the 75% threshold, some questions had responses that reached the threshold in one of the sub-group analysis, Physicists or Clinicians. The presentation of responses by sub-group may be useful for our readers to draw perspectives, and so we have included sub-group results in all our tables.

Our final recommendations, summarized in Table 1, are based on the literature, the consensus views and opinion trends that emerged from the survey and the face-to-face meeting.

Three identical items on patient preparation prior to the MRI scan (Table 3) were included in all PARENCHIMA panels’ questionnaires on clinical renal MRI. To-date, these questions are most relevant for the BOLD MRI panel. A previous systematic review conducted by PARENCHIMA [29] showed that of 20 studies on drug and dietary effects on renal BOLD MRI, 12 studies required their subjects to abstain from food and water for 12 h before MRI, with (8 studies) and without (4 studies) subsequent water loading, 1 study had a shorter fasting period of 2 h before MRI, and 7 studies did not require fasting. The variety of patient preparation in published studies was reflected in the responses to our survey. The respondents did not reach consensus on the need to control subjects’ diet before the scan, although 82% of clinicians were in favor of controlling the diet. Respondents’ comments to the question may explain the lack of consensus:

The respondents reached consensus (87%) that patients should be scanned in a normal (i.e. not restricted to water intake) hydration status, based on studies that show the influence of hydration status on BOLD signal (as mentioned by 8% of respondents). Respondents agreed that hydration status should be controlled, through fasting (4%) or standardized (i.e. all subjects treated similarly in terms of hydration [35]) intake of water (8%).

There was no consensus that sodium intake should be standardized before the scan. 65% of respondents tended to disagree. Even though disagreement among physicists reached consensus threshold (75%), there was a split opinion among clinicians (agree/disagree: 44%/56%). Eight percent of the respondents mentioned their awareness of studies showing that salt intake influences BOLD measurements, with the caveat that the influence of sodium and hydration on the BOLD signal is smaller in the cortex than in the medulla. Modifying salt intake also influences kidney physiology, and may thus confound studies looking at subjects’ typical renal physiology, as commented by another 8% of respondents. 12.5% of respondents motivated their disagreement to this item by the practical difficulties of controlling subjects’ sodium intake, as salt is found in most processed foods. 8% of respondents recommended that controlled sodium intake is feasible for in-patient subjects who can follow a standardized diet during their hospital stay. While standardizing salt intake is not possible or suitable for the study design, monitoring and recording salt intake (e.g. via food diaries) by the subjects can add to any one study’s rigor.

The results of the survey items on data acquisition are summarized in Tables 4 and 7. We sought expert opinion on aspects of BOLD MRI data acquisition protocol including choice of field strength, reduction of B0 inhomogeneities, reducing the effect of intra-renal fat, choice of spatial resolution, magnitude and number of echo times (TE) and control of respiratory motion.

Renal BOLD MRI in humans has been performed at both 1.5 and 3T in almost equal distribution to-date: of 79 renal BOLD MRI human and animal studies included in a previous PARENCHIMA systematic review, 41 were performed at 1.5T, 42 at 3T, and four used magnets of both field strengths [29]. Even though there was consensus (Table 7; 100% of non-abstaining respondents, with 30% abstention rate) that 1.5T provides BOLD data of acceptable quality, there was a general preference for 3T over 1.5T (Table 4) (74% agreement). 3T has the known advantage of greater SNR, increased T₂* weighting and hence higher R₂* values. However, magnetic field inhomogeneities, bulk magnetic susceptibility (BMS) artifacts and fat-water chemical shift are also magnified at 3T [29]. Twenty percent of respondents mentioned these disadvantages of 3T in imaging the kidney near complex abdominal structures like the bowel, when motivating their disagreement or hesitancy to agree with the statement that 3T was superior. For control of respiratory motion, consensus was reached already in Round 1 (Table 7) that breath-holding is the best approach for both native kidneys and transplants (80%), although panelists who disagreed (20%) were familiar with or working on respiratory-triggered sequences.

There was consensus in Round 1 in choosing the smallest receiver bandwidth for the selected maximum TE (and number of echoes) in order to maximize SNR, and most respondents (88%) thought that the number of echoes and spacing between TEs have an effect on data quality (Table 7). It was generally agreed that more acquired TEs can improve robustness of T₂*/R₂* fitting. Eight to sixteen echoes are commonly used in the literature to-date. Expert opinion was evenly split regarding choice of longest TE based on the longest T₂* expected in the kidney (Table 4). While 53% of non-abstaining respondents tended to agree that the longest TE should be equal to the longest T₂* there were concerns that it may be too conservative (e.g. more data can be acquired beyond that value to improve the fit). Conversely, acquiring up to 1.5 times the longest T₂* expected in renal tissue, while preferable to 53% of respondents, raised concerns that the noise floor could be reached in tissues with low oxygenation, that the TR, and hence breath-hold interval, would be increased too much, or that good image quality could not be obtained with some coils at TE = 1.5 times longest T₂* expected in the kidney.

With regards to mitigating (or minimizing) B₀ inhomogeneities, most respondents (65%, Table 4) did not consider the standard abdominal shim performed by the scanner to be sufficient for quality BOLD data. There was overall consensus that a restricted shim should be performed on a volume encompassing both kidneys (76% of all respondents, 91% of clinicians, but only 64% of physicists, agreed), but not on a smaller volume encompassing each kidney separately (79% of all respondents, and 75% of physicists and 78% of clinicians, disagreed). The main concern with shimming on each kidney separately was the time consuming workflow (12.5% of respondents), strongly localized shimming introducing artifacts from outside the shimmed area (8% of respondents), while 1 respondent observed no difference in T₂* maps between the kidneys when shimming for each kidney separately. Respondents using BOLD in the study of kidney transplants used saline bags placed over the skin at the location of the allograft to mitigate susceptibility artifacts from the allograft’s proximity to the body surface and from bowel gas [21, 23]. Respondents acknowledged that localized shimming could not always address B₀ inhomogeneities, which is why there was consensus that acquiring a 3D B₀ map is useful for quality control of BOLD data (80% of non-abstaining respondents, 78% of physicists, and 67% of clinicians agreed). Acquisition of many BOLD datasets with B₀ maps should also foster the development of widely available open-source post-processing software tools that correct the measured R₂* by removing the effect of B₀ inhomogeneity facilitating multicentre studies. Among physicians 67% of non-abstaining respondents agreed on the question of the choice of B₀ map, and there was a high rate of abstaining answers (31–42%), which suggests a low availability or utilization of B₀ mapping sequences and techniques on clinical scanners. However, a dual-echo GRE scheme to generate B₀ maps from the phase difference of the two echoes is now available on all MR vendors. It should be noted that there are no reports in the published literature yet to support these recommendations regarding shimming, which may explain the high rate of abstain responses.

One manifestation of B₀ inhomogeneity is the voxel-size dependence of R₂* estimates [36]. There was evidence to show that slice thickness (usually larger than in-plane voxel size) contributed more to the observed effect [36]. The panelists were asked to consider this evidence [36] when responding to the last four items of the Acquisition portion of the survey (Table 4). While consensus was not achieved on the need for a nearly isotropic resolution, a sizeable portion of non-abstaining respondents (67%) thought isotropic voxels were preferred for imaging an organ with complex structures like the kidney. However, isotropic imaging raised concerns because of the need to match voxels in multi-parametric protocols with other MRI contrasts that are not best suited to an isotropic resolution protocol, such as arterial spin labeling (ASL). There was consensus for using 3 × 3 mm² in-plane resolution (not necessarily isotropic). Similarly consensus was reached for using acceleration factor of 2 to achieve anatomical coverage within a breath-hold while maintaining sufficient SNR.

Studies in the liver demonstrate that the presence of fat can influence T₂*/R₂* measurements [37]. While healthy kidneys do not contain intra-renal fat, in some diseased states such as diabetes [38] and obesity [39], there may be intra-renal fat present. For this and other reasons (delineation of kidney), fat saturation/water excitation is desirable for BOLD MRI acquisitions, but not widely used. The low utilization of these techniques is reflected in the response to the question of whether it is preferable to control for the presence of fat by selection of all echo times with fat-water in phase rather than by performing fat saturation (71% agreement, 42% abstain rate). The statement reached agreement among physicists (78%), but not among clinicians (50/50% split, with 38–54% abstain rate). Because of the lack of consensus, but tendency towards a choice of in-phase fat-water TEs, we recommend performing both fat saturation/suppression and choosing in-phase TEs.

The results of the survey items on data analysis are summarized in Tables 5 and 7. 100% of respondents agreed that the reported value should be R₂* (rather than T₂*) obtained from exponential or log-linear fit of BOLD signal (83%). However, respondents cautioned that the results are not comparable between fitting methods, and that the log-linear fit increases the influence of noise for later echoes with low SNR. A possible solution to correcting for noise at later echoes is to use a weighted log-linear fit. An exponential fit should also be corrected for noise, after identifying the noise floor. Of course, this concern can be mitigated by limiting the maximum TE to be ~ T₂* of tissue of interest.

In Round 1, respondents reached consensus that measuring R₂* values in both the cortex and medulla is equally important (Table 7; 90%). Early renal BOLD MRI research was driven by the low oxygenation and greater sensitivity to changes in oxygen demand in the medulla. Some studies recommended the medulla-to-cortex ratio in R₂* to be a more sensitive marker especially in diabetes, taking advantage of the disparate changes with disease severity in cortex vs. medulla [14]. Increased R₂* in the cortex has been observed in chronic kidney disease (CKD) [19], with either minimal change or actually a reduction in medullary R₂* [15, 40, 41]. The consensus among all respondents (Table 5; 82%) was that a novice BOLD researcher can obtain R₂* measurements in both cortex and medulla using regions of interest (ROIs) drawn in collaboration with a radiologist. R₂* measurements in this case can be based on average signal in the ROI (for protocol optimization), or pixel-based within the ROIs. Ideally, manual ROIs should delineate as much of the cortex and medulla as possible, avoiding lesions, large vessels and fat.

More than with image acquisition, there was greater difficulty in obtaining consensus on advanced, semi-automatic image analysis methods. This is due to the lack of availability and experience with each method beyond the proponent site. Among advanced methods of ROI segmentation, semi-automated segmentation of cortex and medulla using histogram analysis to define masks for cortex and medulla based on T₁ maps [42] nearly reached consensus among 72% of respondents (75% of clinicians, 70% of physicists). Histogram analysis of T₂*/R₂* maps and the 12-layer concentric objects (TLCO) or “onion peel” method [19] generated a 50/50 split in preferences among respondents. There was agreement among physicists (75%), but not among clinicians (50–50% split of opinion) that semi-automated analysis methods, which identify pixels as being “hypoxic” based on a predetermined threshold value for R₂* [28], are not preferred. Concerns regarding implementation, as it is difficult (if not impossible, given the effects of shimming) to decide on a reliable threshold for hypoxia for the condition or patient population studied. Some common concerns expressed by respondents regarding semi-automated segmentation methods were that segmentation software packages are not widely available, have been tested only with particular acquisition protocols (e.g. preference for true coronal to the kidney orientation for the TLCO method [19]), may not work so well in diseased kidneys with poor cortico-medullary differentiation, and have not been validated across sites.

The results of the survey items on BOLD applications and data interpretation are summarized in Tables 6 and 7. In Round 1 (Table 7), consensus (80%) was achieved among respondents that renal BOLD MRI can reflect oxygenation qualitatively, as the confounding factors of perfusion, blood volume fraction (per tissue volume), hematocrit, and shifts of the oxy-Hb dissociation curve preclude it from being truly quantitative. For that reason, respondents consented to the statements that renal BOLD MRI is more suited to detect changes in renal oxygenation within the same kidney rather than for comparing cohorts (88%), and that examining changes in R₂* brought on by a furosemide challenge is useful, although not necessary for all applications (75%).

The consensus statement emerging from Round 1 that BOLD MRI provides a relative measure of oxygen availability, rather than an absolute measure of oxygenation, was supported in Round 2 (Table 6; 87% of respondents). The consensus from Round 1 that BOLD MRI is a potentially prognostic technique (100% non-abstaining respondents) was further improved in Round 2, with 78% of respondents (Table 6) considering BOLD a potentially diagnostic technique. There was notable split among clinicians (92%) and physicists (63%) on the diagnostic value of BOLD. The statement that BOLD MRI is currently only useful as a research tool (Table 6) narrowly missed the consensus threshold with 74% overall. Respondents commented that BOLD can only be diagnostic in combination with other methods, due to the confounding factors, and that its clinical impact has to be better demonstrated beyond a few experienced centers.

Among applications of BOLD MRI, the consensus (Table 6; 86% overall) emerged that renal BOLD MRI is best suited for studying acute responses to physiological or pharmacological maneuvers (e.g. water or furosemide challenge). Fewer respondents considered renal BOLD suitable for evaluation of acute dysfunction in native kidneys or allografts (67% overall, 75% of physicists, 60% of clinicians) or chronic renal dysfunction (53% overall, 44% physicists, 60% clinicians). Although there are studies showing differences between groups in acute dysfunction [43, 44], respondents (8%) commented there are better ways to assess renal dysfunction, and that more work is needed to show clinical relevance of BOLD MRI (8%). The division in opinion on the utility of BOLD in CKD is based on the published studies, which found increased R₂* (lower oxygenation) [14, 45] or decreased R₂* [15] in patients with CKD compared to controls, and did not show a significant correlation of R₂* with estimated GFR or CKD stage [35, 40, 46].

With regards to short and long-term scan-rescan repeatability, the consensus among respondents (85%) was that the literature [47, 48] shows high repeatability for translational research, within the same site, if shimming and physiological conditions (diet, hydration) are controlled. However, 68% of respondents disagreed that multi-center repeatability suitable for clinical trials has been shown in the literature to-date. Respondents commented that standardization of protocols and more accumulation of results from multi-center studies is needed.

Most items that did not reach consensus are related to split opinions in the literature, or limited data available to-date. The fact that our respondents did not reach consensus on superiority of imaging at 3T, standardizing diet, or salt intake reflects the even split among published studies: studies at 1.5T and 3T are equally represented in the literature, and a minority of published studies have standardized salt intake [29]. Split opinion on the longest TE is also due to a lack of systematic comparison studies exploring how feasible it is to image at long TE’s (1× or 1.5 x T₂* of cortex) on different platforms, with respondents relying on their individual experience. More than with image acquisition, there was greatest difficulty in obtaining consensus on advanced, semi-automatic image analysis methods. This is due to the lack of availability and experience with each method beyond the proponent site.

The unresolved issues emerging from this consensus process point to potential avenues for future research. More systematic studies are needed on whether controlling diet beyond a 4-h fast before the scan through standardized diet and salt intake has an effect on renal BOLD MRI. Technical development, validation and dissemination is needed for B₀ mapping and related methods to mitigate B₀ inhomogeneities and bulk susceptibility artifacts. Similarly, respiratory motion control by breath-holding is a limiting factor, creating a trade-off between the number of echo times or slices acquired, which can be addressed by further development of fast free-breathing BOLD MRI sequences or respiratory motion correction methods. Respiratory motion control beyond breath-holding is essential for achieving whole kidney coverage in BOLD MRI acquisitions.

On the data analysis side, the survey responses underline the need for dedicated renal BOLD MRI post-processing software to facilitate semi-automated image analysis, with validation and dissemination of the software beyond the most experienced sites. The development and validation of (semi-)automatic segmentations methods for renal BOLD MRI data should be encouraged, especially in patient populations with poor cortico-medullary differentiation, where manual selection of cortical and medullary ROIs can be biased.

There is a future role for artificial neural networks/ artificial intelligence (AI) in solving cortex and medulla segmentation problems on BOLD MRI data acquired in patients with renal disease. To date, there are no published studies on AI in renal BOLD MRI, but artificial neural networks have been used in the brain for quantitative analysis of BOLD MRI and quantitative susceptibility mapping data [49]. AI methods have been used for automatic segmentation of the cortex and medulla for renal volumetry based on T₁-weighted imaging, with [50] and without gadolinium contrast enhancement [51‐53]. The ROIs obtained by automatic segmentation of T₁-weighted images can inform segmentation of the cortex and medulla in a multiparametric MRI protocol. We plan to include a literature review as the field develops, and survey questions on AI in renal BOLD MRI in future iterations of the recommendations. Use of AI requires large training sets of manually annotated, high-quality renal BOLD MRI data, with robust clinical validation and correlation with patient-centric outcomes [54]. Increased quality and standardization in data acquisition is thus needed for the large-scale deployment of AI in renal BOLD MRI. Multicenter studies with harmonized scanning protocols, uniform semi-automatic data analysis, and the development of a phantom with reference cortical and medullary R₂* values are high priority for the qualification of BOLD MRI R₂* as a renal biomarker. Additional challenges are also likely to emerge with better understanding of the kidney’s complex physiology and pathophysiology, such as the degree to which renal functional markers are influenced by the huge range of drugs used to treat renal and systemic diseases, as well as how these measured renal parameters change in multi-morbidity states.