Hyperpolarized 83Kr magnetic resonance imaging of alveolar degradation in a rat model of emphysema

Hyperpolarized 83Kr surface quadrupolar relaxation (SQUARE) generates MRI contrast that was previously shown to correlate with surface-to-volume ratios in porous model surface systems. The underlying physics of SQUARE contrast is conceptually different from any other current MRI methodology as the method uses the nuclear electric properties of the spin I = 9/2 isotope 83Kr. To explore the usage of this non-radioactive isotope for pulmonary pathophysiology, MRI SQUARE contrast was acquired in excised rat lungs obtained from an elastase-induced model of emphysema. A significant 83Kr T1 relaxation time increase in the SQUARE contrast was found in the elastase-treated lungs compared with the baseline data from control lungs. The SQUARE contrast suggests a reduction in pulmonary surface-to-volume ratio in the emphysema model that was validated by histology. The finding supports usage of 83Kr SQUARE as a new biomarker for surface-to-volume ratio changes in emphysema.


Introduction
Hyperpolarized krypton-83 (hp 83 Kr) enables MRI contrast that is indicative of surface composition [1,2] and the surface-to-volume ratio (S/V) [3] in porous media. The T 1 -weighted MRI contrast is generated through surface quadrupolar relaxation (SQUARE) that causes S/V-dependent reduction in the hp 83 Kr MR signal intensity as sketched in figure 1. Pulmonary SQUARE MRI contrast between major pulmonary airways and the alveolar regions has recently been demonstrated in excised rat lungs [4]. In this publication, the potential of hp 83 Kr SQUARE contrast for lung pathophysiology is evaluated using ex vivo MRI of an established rat model of emphysema. The emphysema model was selected for this proof of concept study because of the well-known deterioration of the alveolar surface that is expected to cause increased SQUARE T 1 times. Following the MRI acquisition, the alveolar surface deterioration can be quantified through histology.
Emphysema is a component of chronic obstructive pulmonary disease (COPD), the fourth leading cause of death worldwide [5]. The development of emphysema is mainly linked to cigarette smoking with a smaller proportion attributed to pollution, occupational exposure or intrinsic factors [6]. Over time, there is significant alveolar destruction with the resulting reduction in surface area for gas exchange with an accompanied loss of lung elasticity. Unfortunately, current routine investigations such as lung function tests often fail to diagnose the disease until the later stages [7] once a significant amount of damage has been done. There is therefore a need for new biomarkers to detect the disease in the early stages and to help separate COPD phenotypes [8]. Investigations of COPD and emphysema, in particular, have developed over the past decade with human studies using both computed tomography (CT) [9] and magnetic resonance imaging (MRI)-based techniques [10,11]. Hyperpolarized noble gas MRI [12][13][14] using 3 He and 129 Xe is able to provide measurements of ventilation and is able to delineate poorly ventilated and non-functioning lung regions [11,12]. Recently, visualization of delayed collateral ventilation into lung regions has been performed [15], providing information complementary to CT-based techniques. Furthermore, hp 3 He and hp 129 Xe provide measurements of the lung microstructure through the apparent diffusion coefficient (ADC) [16][17][18][19][20][21][22][23][24]. Using animal models of emphysema, the distinction between healthy and emphysematous tissue was possible through fractional ventilation generated hp 3 He MRI contrast before an increase in alveolar diameter could be determined through histology [25,26]. Unlike fractional ventilation that is decreased in the disease model, ADC was found to be increased in long-term disease models developed over a six-month period [21]. A very promising new pulmonary MRI contrast can be obtained by probing the dissolved phase of hp 129 Xe. A host of innovative new techniques, such as xenon polarization transfer contrast (XTC) [23,[27][28][29][30], xenon alveolar capillary transfer (XACT) [31] and chemical shift saturation recovery spectroscopy (CSSR) [29,32,33] enable selective detection of gas phase xenon, tissue and plasma (TP) dissolved xenon and xenon interacting with red blood cells (RBC). For example, Dregely et al. [23] found a strong correlation between the XTC-based parameter (MXTC-F) and CT data. Using CSSR spectroscopy, Patz et al. [33] found a dramatic decrease in S/V in emphysema patients, while alveolar septal thickness and capillary transit time was not affected. In general, the findings suggest that a sensitive probe for S/V changes is the key to early emphysema diagnosis.
This work sets forth the verification of a new type of hyperpolarized noble gas modality for pulmonary studies, i.e. hp 83 Kr SQUARE MRI contrast [4], which may provide a significant addition to existing methodology. Previous work exploring the underlying conceptual physics found SQUARE to be the dominant cause of 83 Kr T 1 relaxation observed in the gas phase in high S/V porous media such as lungs. See figure 1 and Material and methods section for an explanation of the SQUARE concept. The purpose of this work is to determine whether the SQUARE effect is sensitive enough to serve as a probe for disease-related lung physiological changes.
The potential significance of SQUARE MRI contrast is that it may enable a novel type of biomarker for pulmonary pathophysiology through a fundamentally different physical effect compared with those used in other pulmonary diagnostic techniques such as hp 3 Figure 1. Illustration of the concept of hp 83 Kr SQUARE contrast. The 83 Kr atoms are depicted with the 10 possible spin states of this nuclear spin I ¼ 9/2 isotope and with a non-spherical nucleus that possesses a nuclear electric quadrupolar moment. Hyperpolarized 83 Kr atoms, drawn in blue, are depicted as occupying the lowest spin state (i.e. corresponding to the highest MRI signal intensity). However, longitudinal T 1 relaxation will cause the spins to assume other states (atoms indicated in red), thus reducing the MRI signal intensity. Quadrupolar relaxation occurs when the electron cloud of 83 Kr is forced out of the spherical shape.
(a) In the bulk gas phase a non-spherical symmetry is caused by collisions (i) leading to the gas phase signal decay within minutes. (b) When surfaces are present, surface adsorption (ii) and possibly dissolution into deeper surface regions (iii) will lead to surface quadrupolar relaxation (SQUARE). Rapid exchange causes transfer of the SQUARE effect into the gas phase where the rapid signal decay is observed with typical T 1 times approximately 1 s in rat lungs. See Material and methods section for further details.
of the 'surface probing' is 1-2 orders of magnitude longer than that of ADC measurements. The nature of the contrast generation may provide a methodology sensitive not only to S/V but also to the chemical composition of the surface. Furthermore, in model surfaces quadrupolar noble gas (i.e. 131 Xe) relaxation was affected by microscopic surface fine structure [34,35]. Similarly, 83 Kr SQUARE is likely to be sensitive to surface corrugation and may, therefore, provide different S/V values compared with dissolved xenon-based techniques [33]. Informed by the surface studies of previous models [1][2][3]36,37], the underlying hypothesis of this publication is that hp 83 Kr SQUARE MRI contrast can serve as a biomarker for the alveolar S/V reduction caused by emphysema. Confirmation of the S/V hypothesis in a preclinical small animal model is a crucial milestone for the development of hp 83 Kr MRI. The intratracheal elastase exposure of rat lungs to porcine pancreatic elastase (PPE) generates an established model of emphysema whereby initial inflammation is produced with the subsequent development of airspace enlargement and destruction [38 -41]. The experimentally and regulatory less demanding ex vivo set-up [14,42] was used for the MRI measurements in this proof of concept work to set the basis for future in vivo preclinical and clinical studies. Figure 1 illustrates the mechanism of SQUARE with the nuclear spin I ¼ 9/2 isotope 83 Kr. Atomic nuclei are positively charged electric monopoles, however, the nucleus of any spin I . 1/2 isotopes is non-spherical leading to a non-uniform nuclear electric charge distribution. The resulting nuclear electric quadrupole moment can interact with the surrounding electrons if the noble gas atom is 'distorted'-i.e. if the electronic cloud assumes a non-spherical symmetry. In the bulk gas phase, far away from surfaces (figure 1a), collisions events with other atoms cause rapidly fluctuating electron cloud distortions that results in 83 Kr quadrupolar relaxation with T 1 times of several minutes at ambient pressure. SQUARE (figure 1b) occurs when surface adsorption and possibly dissolution into deeper surface regions takes place. SQUARE can typically not be observed directly owing to strong line broadening and the limited number of atoms at the surface at any given time. However, the SQUARE effect is transferred into the gas phase through rapid exchange and, depending on the S/V ratio, can tremendously accelerate the decay of the hp gas phase signal. SQUARE contrast is, therefore, potentially sensitive to S/V, surface composition and surface temperature. SQUARE caused by high S/V alveolar region of rat lungs reduces the gas phase T 1 time to approximately 1 s leading to the rapid hp 83 Kr signal decay in the series of small flip angle spectra in figure 1b.

83 Kr Spin exchange optical pumping, compression and transfer
Hyperpolarized 83 Kr was produced in batch mode by SEOP as described previously described in detail [43]. All MRI was performed using enriched 83 Kr (99.925% 83 Kr, CHEMGAS, Boulogne, France) to improve the available signal intensity. A 15% krypton 85% N 2 (99.999% purity, Air Liquide, Coleshill, UK) mixture was used to reduce the consumption of expensive isotopically enriched 83 Kr. SEOP build-up times of 12 min, corresponding to greater than 92% of the steady-state polarization were used to reduce the experimental duration. The hyperpolarized gas extraction unit described previously [4,44] was used to allow for below ambient pressure SEOP [4,43] performed at 55 -65 kPa [43]. Overall, the method produced a 83 Kr nuclear spin polarization of P ¼ 16 -17% after accounting for depolarization occurring in the gas extraction process [44]. An approximate volume of 12 -16 ml of the hp gas mixture (1 : 6.7 Kr:N 2 ) was obtained for lung imaging every 12 min. Since no viable method currently exist to separate hp 83 Kr from the mixture, it is instructional to report the apparent spin polarization of approximately P app ¼ 2.5%. The apparent polarization is the nuclear spin polarization P times the fraction of krypton in the hp gas mixture [43]. A discussion of the nuclear spin polarization P for isotopes with nuclear spin I . 1/2 can be found in [45].

Hyperpolarized gas inhalation
The lungs were suspended in a 5% glucose solution (weight/ volume) in the ventilation chamber as described in previous work [14,42]. The chamber was then placed in the centre of the superconducting magnet bore with the temperature kept constant at 295 K throughout the experiments. Active inflation of the lung was accomplished by pulling to a ventilation syringe volume (V s ) of 8 ml. Corresponding inhaled volumes (V i ) were measured separately using the water displacement technique on gas exhalation [14] and are shown in table 1. To limit gas trapping ( particularly notable in the PPE-treated lungs) the ex vivo lungs were deflated over 30 -60 s from V s ¼ 8 ml to maximum exhalation (V s ¼ 0 ml) as has been reported elsewhere [46,47] before hp 83 Kr inhalation.

Magnetic resonance imaging protocol
MRI experiments were performed using a vertical bore 9.4 T Bruker Avance III microimaging system (Bruker Corporation, Billerica, MA, USA) with a standard Bruker double saddle coil tuned to 83 Kr resonance frequency of 15.40 MHz. The internal diameter (ID) of the coil was 30 mm. Coronal images were acquired into 64 Â 32 matrices using a variable flip angle (VFA) FLASH protocol (TE ¼ 1.8 ms, TR ¼ 12.6 ms) [48]. Rectangular RF pulses of constant duration of 0.3 ms and variable power levels were used in all experiments. The imaging protocol had a total acquisition time 0.405 s to minimize the effects of unwanted T 1 decay during acquisition. To obtain T 1 -weighted images [4], each imaging sequence was started after a programmed time delay, t d , post inhalation. Typically, t d ¼ 0.2 s, 0.7 s, 1.2 s, 1.7 s and 2.2 s were used, although some of the T 1 maps were calculated from a series of four images with t d ¼ 0.5 s, 1.0 s, 1.5 s and 2.0 s. The inhalation itself was accomplished manually by reducing the pressure in the artificial pleural cavity using the ventilation syringe [14,42]. Slight alternations in the timing of the manual inhalation procedure (approx.+0.2 s) were deemed acceptable. In all imaging experiments each individual image was acquired from a single inhalation cycle using one VFA FLASH acquisition and no signal averaging. The resulting FOV was 50.9 Â 40.7 mm 2 .

Image reconstruction and T 1 analysis
The raw 32 Â 32 datasets were apodized using sine-bell squared function and zero filled to 64 points in each spatial domain before Fourier transformation. Final image resolution was 0.795 Â 0.635 mm 2 in the frequency encoding (longitudinal) and in the phase encoding (transverse) directions, respectively. These final 64 Â 64 magnitude images were exported to IGOR Pro (v. 6.01, Wavemetrics, Lake Oswego, OR, USA) for T 1 analysis.
The T 1 datasets were created for each series of images by combining the images acquired at the individual time delays t d into a three-dimensional dataset where the first image in rsif.royalsocietypublishing.org J. R. Soc. Interface 12: 20150192 the set corresponded to the image acquired at the smallest time delay, t d as described in [4]. The T 1 values outside of the 0 s , T 1 6 s range were rejected as physically not meaningful. The T 1 values within the range of 0 s , T 1 6 s but located outside the lung contour region, or the region of interest (ROI), were also discarded. The ROI was determined from the first image in each T 1 set. These final T 1 maps were used to produce T 1 histograms for each lung reported in this study.
The T 1 data collected from the ROI in each T 1 map were binned into 200 intervals with 0.03 s increment. The resulting histograms were analysed using build-in Multipeak 1.4 fitting procedure and automated peak picking. A bi-modal Gaussian distribution of T 1 values in the lung was assumed because of the distinct alveolar and conducting airway compartments present in the lung. The results of the analysis returned the most probable (expected value, EV) relaxation time, T EV 1 , its probability and distribution measured as full width, half maximum, FWHM(T 1 ) for each of the two Gaussian components. followed by 20 ml 21 Dublecco's phosphate buffer solution (D-PBS, Sigma-Aldrich Ltd, Gillingham, UK) to remove residual blood from the pulmonary circulation. The heart and lungs were subsequently removed en masse. Lungs for ex vivo hp 83 Kr imaging had a polytetrafluorethylene (PTFE) adapter tube inserted into the trachea 5 -10 mm above the carina and sutured into place. The heart and lungs were then suspended in 5% glucose solution (weight/volume) (Baxter Healthcare Ltd, Thetford, UK) with the trachea pointing downwards as has previously reported [14,42]. The ex vivo lungs were repeatedly inflated with 8 -10 ml of room air to check for gas leaks either from the suture around the trachea or the lungs themselves. The lungs were chilled to 278 K for transportation to the imaging facility during which time they were repeatedly inflated with 8 -10 ml of air at 30 -60 min intervals. Time from lung harvest to the start of imaging was no more than 8 h for each lung.

Model characterization: alveolar cross-section measurements
PPE-induced changes in air space were measured in both a satellite group of rats and on the ex vivo lungs used for hp 83 Kr MRI. Lungs from both groups were similarly prepared by inflating to forced vital capacity (25 cm H 2 O) with 5% formalin -saline solution (Sigma-Aldrich Ltd, Gillingham, UK). The trachea was then tied off and the lungs and heart removed en bloc and placed in formalin. After at least 24 h in formalin, an experienced histologist processed the lungs. Sections were cut from the wax embedded samples using a microtome and stained with H&E. Experimental emphysema was assessed by measuring average air space area using a method describe by Belloni et al. [50] and used previously [40]. Briefly, computer-assisted image analysis was performed using an Olympus BX40 microscope and Zeiss image-processing software (Imaging Associates, Bicester, UK). Using colour thresholding techniques, the total alveolar air space area and the number of air spaces was measured from five random fields per sample. From these figures the mean air space area was calculated. Any fields Table 1. Demographic data from satellite subjects (histology only) and those used for hp 83 Kr imaging (with subsequent histology). Summary of rat weights, whole lung mean alveolar area + standard deviation of the mean, inhaled volumes (V i ) + standard deviation corresponding to inflation (syringe) volume V s ¼ 8 ml with associated inhalation pressures + standard deviation. No values for V i were determined in the histology groups. Values omitted were not measured.  figure 1). In any case however, the 83 Kr relaxation time is expected to be sensitive to the S/V ratio and the purpose of this study is to investigate whether SQUARE can serve as an indicator for the emphysema model. Although other factors, such as chemical composition affect SQUARE [3] the expected strong dependence on S/V ratios is the likely cause for regions with lower S/V, such as the major airways in the control lung, to experience a slower T 1 relaxation and thus a slower depolarization rate. These areas remain 'bright' in the MR images while areas with higher S/V lead to 'dark' regions of faster depolarization. The T 1 map calculated from the delay time t d -dependent signal decay leads to the actual SQUARE contrast images shown in figure 2b for the control lung and figure 2d for the elastase-treated lung. Prolonged relaxation times were found for the major airways compared with the alveolar region in the control lung in figure 2b (see also ref. [4]). PPE-treated lungs, serving as a model for emphysema, are imaged for the first time with hp 83 Kr in this work. The hp 83 Kr SQUARE T 1 map of the lung shown in figure 2d exhibits increased T 1 values for the alveolar region compared with control lungs. This is reflected in the SQUARE T 1 map that displays an elevated level of green colour compared with figure 2b (far right). , is displayed by the green dashed Gaussian curve. The sum of both components results in the black solid line that very closely resembles the histograms, suggesting that bimodal fitting is a good approach for extraction of parameters that are characteristic for the SQUARE behaviour of a lung. Figure 4 shows the SQUARE T 1 maps and histograms for PPE-treated lungs. SQUARE T 1 maps and histograms display marked differences between PPE-treated and control lungs as can also be noted by the characteristic T 1 times from the bimodal fitting listed in table 2. The combined histograms for all three control lungs are shown together with the combined histograms of the four elastase-treated lungs in figure 5. Figure 6 displays a fifth lung (EL.5) from the PPE group that exhibits SQUARE characteristics very similar to that of control lungs. However, table 1 shows that the averaged MAA is fairly low with MAA ¼ 3.6 Â 10 4 mm and thus is similar to that of the control group.

Discussion
Two of the control lungs in figure 3 display clear ventilation heterogeneity as parts of the lungs remain dark in the hp 83 Kr MR images. The cause of the heterogeneity is unknown but the lungs were transported over a 3-h period, subsequent to excision, between the location of animal holding facility at Imperial College and the hp MRI facility at Nottingham. Nevertheless, the analysis of the SQUARE T 1 maps and histograms of all three control lungs led to comparable results with similar bimodal T 1 distribution. Figure 5 depicts the combined T 1 frequencies for the control group and for the elastase-treated animal lungs. The characteristic T 1 values from bimodal fitting are listed in table 2 and show little variation within the two groups (i.e. control and PPE group). The pronounced increase in the T EV 1 times for the PPE group in figure 5 and table 2 shows that the T EV 1 data can serve as an indicator for the development of the symptoms in the emphysema model. Presumably, the T EV 1 data are correlated to MAA. This correlation is expected from previous work with model surfaces but requires further corroboration in future studies. In any case however, the FWHM(T 1 ) data are not associated with the MAA differences between the two groups. This view is further supported by the box diagrams in figure 7, which show no overlap in T EV 1 data between the two groups. However, the associated FWHM(T 1 ) data almost completely overlap and are therefore unlikely to be useful as a biomarker for changes in MAA associated with the disease model. Figure 7 also provides the results of statistical significance testing using Student's t-test at a critical significance level of A correlation between MAA and SQUARE would be caused by the dependence on lung S/V ratio. A decreasing S/V is expected from increasing MAA in the emphysema model [21,26] and it is known from previous work with model surfaces [3] that T 1 / V/S. In this work, the T 1 increase was not quantitatively correlated with the increase in MAA but a statistically significant increase in T EV(f) 1 and T EV(sl) 1 was observed for the emphysema model with increased MAA in histology compared with the control group. The bimodal appearance of the histogram could be caused by the presence of high S/V in the alveolar region and lower S/V within the airways. The alveoli may cause the fast relaxing mode described by T EV(f) 1 with a narrow FWHM(T (f) 1 ) as the alveolar region will be fairly homogeneous. The airways, on the other hand, may contribute to the slow T EV(sl) 1 mode with a broad distribution and hence a larger FWHM(T (sl) 1 ) due to the higher variability in S/V within the various airway generations. Note the relative contribution (in pixel count) between the two modes may not be quantitative due to the different extent of depolarization during the initial 0.2 s (fixed) delay and image acquisition. If the slow mode is indeed (at least partially) caused by the airways, the increase in T EV(sl) 1 in the elastase-treated rat lungs would suggest that the S/V in the airways is affected by the disease model. Although this point was not further investigated, nitrogen chaser hp 83 Kr MRI could be used for future exploration where the airways are largely purged with a small volume of N 2 at the end of the inhalation [42].
The fifth lung (EL.5) from the PPE-treated group was excluded from the statistical analysis because the MAA increase developed asymmetrically in the left lung lobe only, Table 2. Characteristic T 1 times from bimodal fitting of the histograms of all lungs used in this work.  The SQUARE T 1 map and the corresponding histogram are indeed very similar to that of the control lungs, however, this also includes the T 1 values found in the left lobe. One would expect that the unexposed part of the lung to shows control lung behaviour, as indeed observed, but one would also expect very long T 1 times from the left lobe. However, the left lobe may have been damaged too excessively and some regions may, therefore, no longer be ventilated. These areas will not be able to contribute to the SQUARE T 1 map. Although non-ventilated areas should appear as 'dark' regions in the hp 83 Kr MR images, these zones may be masked by MR signals from unaffected regions that contribute to the non-slice selective images.
For this work, excised lungs have been used. Two of the control lungs exhibited ventilation defects that may have been   rsif.royalsocietypublishing.org J. R. Soc. Interface 12: 20150192 caused by the excision procedure and long transport period between the laboratories. Although this situation is non-ideal, the experiments demonstrate that the SQUARE T 1 contrast was little affected by any damage arising from lung excision, ex vivo transport, and ex vivo experiments. The MAA, obtained after MR image acquisition, were in the expected range for PPE and control group, except for EL.5 as discussed above. This work demonstrated that hp 83 Kr SQUARE can serve as a biomarker for the elastase model of emphysema, in all likelihood because SQUARE senses changes in S/V. If the absence of harmful X-ray radiation is a strong motivation to explore the various pulmonary MRI techniques, a further potential advantage arises with hp 83 Kr: the SQUARE effect can be used as a pulmonary biomarker without the presence of pulsed magnetic field gradients (PFGs). Unlike ADC measurements where PFGs are required to generate the MRI contrast, SQUARE is solely caused by a relaxation rooted physical effect. Although PFGs are still needed for the MR image generation, hp 83 Kr SQUARE may have a potential as a global lung surface biomarker for pulmonary screening without spatial resolution. Since the effect of relaxation can be measured remotely [37], SQUARE may be observed without an MRI scanner if the effect is remotely detected upon exhalation into a small bench-top NMR device.
The study presented here was focused on providing a first demonstration of hp 83 Kr SQUARE usage as an MRI contrast and biomarker for pulmonary pathophysiology in an animal model. Remarkably, using completely different physical concepts from those described here, krypton gas usage has been previously reported for contrast generation in dual energy computed tomography (DECT) [52]. The high-volume inhalation of a mixture of 80% krypton and 20% O 2 did not lead to adverse effects in COPD patients. Compared with CT, SQUARE MRI is expected to require lower krypton quantities for usable contrast generation. Finally, molecular oxygen does not dramatically accelerate 83 Kr T 1 relaxation and SQUARE measurements in rodent lungs [42] were unaffected by the presence of up to 40% O 2 .

Conclusion
Informed by the previous S/V dependence of SQUARE in model surfaces, the aim of this work was to demonstrate the ability of hp 83 Kr SQUARE MRI to serve as a biomarker for the elastase model of emphysema. It was demonstrated that two characteristic T EV 1 times, obtained from bimodal fitting of the histograms, enable statistically significant distinction between emphysema model and control lung. Beyond statistics, the difference between the control group and emphysema model can also be identified from visual inspection of the 83 Kr SQUARE images shown in figures 3 and 4. The use of the simpler ex vivo model has allowed for rapid confirmation of the imaging technique despite some ventilation defects in the control group. Hp 83 Kr SQUARE may serve as a potential biomarker for pulmonary disease-related S/V ratio changes. Future quantification of the effect and comparison with other hp noble gas modalities will provide further evaluation of this technique. Because neither the application of magnetic field gradients nor the detection within high magnetic fields is required for SQUARE measurements, this new biomarker should also be explored for potential pulmonary mass screening using small bench-top devices.
Competing Interests. We declare we have no competing interests. data. In contrast, the distribution of T 1 values for the two modes of a histogram (characterized by FWHM(T 1 )) significantly overlaps between control and PPE group, indicating that no significant statistical difference to be expected from the T 1 spread. However, the variation in T EV 1 , in particular T EV(sl) 1 , between the individual animals is more pronounced in the PPE group than in the control animals as visible in the added data in figure 5. rsif.royalsocietypublishing.org J. R. Soc. Interface 12: 20150192