In vitro and in vivo comparison of a tailored magnetic particle imaging blood pool tracer with Resovist

Both tracers, ferucarbotran (Resovist, Bayer Pharma AG) and LS-008 (LodeSpin, Seattle, USA) have superparamagnetic iron oxide cores. Resovist is polydisperse and coated with negatively charged carboxydextran (Lawaczeck et al 1997). LS-008 consists of iron oxide cores with a median diameter of 25 nm and a monodisperse size distribution ($\sigma \approx$ 0.05, where $\sigma$ is the standard deviation of the number-weighted lognormal distribution function). The iron oxide cores in LS-008 were synthesized by the thermal decomposition of iron(III)-oleate with oleic acid in 1-octadecene. Details of the core synthesis can be found in Kemp et al (2016). The as-synthesized cores were hydrophobic and required an additional phase transfer step to disperse in aqueous solvents. Oleic acid-coated nanoparticle cores were encapsulated with an amphiphilic poly(maleic anhydride-alt-1-octadecene) (PMAO, M_n  =  30–50 kDa), which was grafted with methoxy-polyethylene glycol-amine (m-PEG-NH₂, M_n  =  20 kDa) conjugated at 18.8% of the number of carboxylic acids in PMAO (Khandhar et al 2017). The final composition of LS-008 consisted of monodisperse PEGylated iron oxide cores at a concentration of 87 mM Fe in 1x-phosphate buffered saline. Surface characterization, using dynamc light scattering (Zetasize Nano ZS, Malvern Instruments), showed a z-average hydrodynamic diameter of 80 nm and a mostly neutral surface charge.

MPS measurements can proof the tracers' potential for the imaging performance in MPI. All MPS and MPI measurements were performed with a preclinical scanner (Bruker/Philips, Ettlingen/Hamburg, Germany) which provides three orthogonal adjustable drive field channels. While the amplitude and strength can be adjusted the frequencies of the drive fields are fixed and given by f_x  =  24 509.8 Hz, f_y  =  26 041.7 Hz, f_z  =  25 252.5 Hz. This leads to a period of the Lissajous trajectory of 21.53 ms covering a 3D volume. In principle, any MPI scanner can be used as MPS when the gradients of the selection field is chosen to be so small, that the field free point is larger than the sample. In this case the MPI signal is integrated over the entire sample. The gradient strength was set to 0.05 T m⁻¹ for the MPS measurements since zero gradient lead to substantial background signal distorting the MPS spectrum. The drive field in x direction was used for the MPS measurements with an amplitude of 14 mT. 50 µl samples of LS-008 and of Resovist with each of a concentration of 45 mM were filled in 0.2 ml PCR tubes (Eppendorf AG, Hamburg, Germany). Measurements were carried out with 100 averages and 1000 repetitions.

For image reconstruction system functions reflecting properties of the tracer and the scanner are mandatory. Measurements of system functions were performed separately for LS-008 and Resovist using a robot and a point sample filled with 4 µl of 87 mM iron which covered a volume of 2  ×  2  ×  1 mm³. Resovist was diluted to exhibit the same iron concentration as LS-008. For both tracers the same parameters were used: drive field amplitude 14 mT, gradient field strength 1.5 T m⁻¹, field of view of 37  ×  37  ×  185 mm³, bandwidth 1.25 MHz, and 600 averages. The resulting system matrix consisted of 25  ×  25  ×  25 image voxels covering a volume of 50  ×  50  ×  25 mm³. After 25 robot positions a background noise measurement was performed resulting in an overall scan time of 64 h each.

Similar to MPS a MPI scanner gathers spectral information while performing a system function measurement. While during an MPS experiment the particles are always excited along their easy axis, the situation is considerably more complex for multidimensional excitations as applied during an imaging experiment. Therefore, it is essential to compare and validate a potential signal increase observed in MPS by analyzing the behavior in the multidimensional system function. For simplification of the analysis procedure we restricted our view to the x-channel and to the central robot position corresponding to the signal within the FFP. Background measurements performed without a tracer sample are used for calculating SNR values for each frequency (Franke et al 2016). The calculation of the SNR values was performed by the scanners software. During reconstruction generally a SNR threshold is applied to cancel out frequencies dominated by noise (Rahmer et al 2012).

$$\begin{eqnarray}&&\text{filtered}\,\text{frequencies}=\left\{\,{{f}_{i=1\ldots N}}~|~\forall {{f}_{i}}\in \left[{{f}_{\min}},{{f}_{\max}}\right]:~\text{SNR}\left({{f}_{i}}\right)\geqslant \text{SN}{{\text{R}}_{\text{threshold}}}\right\}\end{eqnarray}$$

The group of filtered frequencies contains all remaining frequencies after thresholding and N  =  |Filtered| is its number. Once filtered it is straight forward to sum up all SNR values to build an integrated measure (ISNR) which can picture the overall performance.

$$\begin{eqnarray}&&\text{ISNR}\left({{f}_{\min}},~{{f}_{\max}}\right){{|}_{\text{SN}{{\text{R}}_{\text{threshold}}}}}=\underset{i=1}{\overset{N}{\sum}}\,\text{SNR}\left({{f}_{i}}\right)\end{eqnarray}$$

To find a suitable threshold we analyzed the frequencies between 1 MHz and 1.25 MHz where amplitudes are low and mainly noise is present. Starting with a threshold of one, letting pass all frequencies, we successively increasing the threshold stopping when ISNR(1 MHz, 1.25 MHz)/250 MHz  <  10⁻² MHz⁻¹ was reached. Next the number of the remaining frequencies N was counted and the frequency f_N determined, defining the highest frequency passing the SNR threshold. Finally, the ratio ISNR_LS-008 (0 kHz, 1.25 MHz)/ISNR_Resovist (0 kHz, 1.25 MHz) over the complete spectra was calculated. This ratio can be directly interpreted as sensitivity amplification when using LS-008 instead of Resovist.

We judged the image quality of the resolution phantoms measured with various tracer concentrations. The application of a resolution phantom is essential since there is a direct but nonlinear link between spatial resolution and sensitivity, which makes theoretical predictions challenging (Ludwig et al 2013). The phantom consisted of a block of Polytetrafluorethylen with five holes of five different sizes of 5, 4, 3, 2, and 1 mm. The distances between the holes were in the order of 5, 4, 3, and 2 mm. For each tracer, LS-008 and Resovist, we filled six phantoms with iron concentrations of 87, 29, 9.7, 3.2, 1.1, and 0.4 mM. To avoid vaporization the samples were sealed with a foil. Positioning was performed by the same robot as used for system function measurements.

Images were reconstructed offline with an in house written reconstruction framework, written in Julia (Bezanson et al 2015). It is based on the iterative reconstruction scheme introduced in Knopp et al (2010) that solves a first order Tikhonov regularized least-squares functional. This is also known as the Kaczmarz algorithm. In order to obtain the best reconstruction result in terms of spatial resolution and noise the regularization parameter has to be individually optimized for each particle concentration. In general, a worse SNR in the measurement data has to be compensated for by a larger regularization parameter. To minimize noise enhancement, especially for the low concentration phantoms, we applied only one iteration step. The applied signal-to-noise ratio threshold (Rahmer et al 2012) was 1.4 as found by analyzing the system functions. Additionally, as we found substantially background noise below 80 kHz, a high pass filter was applied limiting the bandwidth to 0.08–1.25 MHz. That meant also that the third harmonic and all corresponding mixing frequencies were neglected since it was not possible to reliably subtract the background signal in a background subtraction step (Them et al 2016). 2000 averages were applied to improve the SNR.

To study the effect of the regularization parameter λ on the image quality, λ was varied in image reconstruction with steps of 2.5, 0.5, 10⁻¹, 10⁻², 10⁻³, 10⁻⁴, 10⁻⁵, and 10⁻⁶. A standardized region-of-interest (ROI) analysis was performed. Hereby, one ROI_tracer was placed over the area of holes filled with tracer material covering 17  ×  9 pixels and the maximal signal values was determined. Next the mean noise floor was determined by surrounding the ROI_tracer by a second ROI_noise (two times 17  ×  4 pixels) excluding the ROI_tracer. The maximal SNR value was calculated by SNR_max  =  max(ROI_tracer)/mean(ROI_noise) for each λ and concentration.

From the SNR_max analysis one regularization parameter was chosen to be a compromise for the different concentrations. These images were rated by an observer for image quality in concern of spatial resolution and sensitivity. Spatial resolution was judged by the delineation of the holes filled with tracer material. Sensitivity was judged by counting the number of holes that were visible in the MPI looking in each concentration step. When the reconstruction failed to resolve the filled holes the corresponding concentration was determined.

In vivo measurements in six healthy FVB mice were approved by the local committee on animal protection (Behörde für Gesundheit und Verbraucherschutz, Freie und Hansestadt Hamburg, Nr. 42/14). The six FVB mice (age: 8 weeks; average weight: 25 g) were divided in two equally sized groups of three animal each. The first group received 60 µl LS-008 with 87 mM iron, the second diluted Resovist of identical concentration and volume. Tracer injections were performed with a syringe pump (AL1000—220Z, World Precision Instruments) during a dynamic MPI scan.

The workflow was performed as previously published (Kaul et al 2015). Mice were anesthetized with 1–2% Isofluran at a flow of 0.5 l min⁻¹ (Vapor 19.3, Dräger, Lübeck, Germany and Tec 7 Vaporizer, GE Healthcare, Chalfont St Giles, UK). Respiration was monitored throughout the experiments. A catheter (inner tube diameter 0.28 mm, Portex, Smiths Medical International Ltd, USA) was fixed in the tail vein for the injection of the tracer. Fiducal markers were placed on the mouse (Franke et al 2016).

First, MRI was performed for anatomic referencing. The MRI protocol consisted of a survey scan and three respiratory triggered T2-weighted scans covering the chest and abdomen in coronal, sagittal, and transverse orientations. The following scan parameters were used for the 2D turbo spin echo sequence: field of view (FOV) 32 mm, matrix 256  ×  256, slices 28, thickness 0.8 mm (no gap), TR 1100 ms (triggered on every respiratory cycle at 40 cycles min⁻¹), TE 28 ms, turbo factor 8, NSA 3 with a scanning time of 8 min each.

Then the mice positioned in the mouse bench were transferred to the MPI scanner. Fiducial markers and an online reconstruction procedure (Knopp and Hofmann 2016) were used to ensure that heart and liver were positioned into the central part of the field of view (Werner et al 2016). All MPI scans were performed with the same hardware settings as the in vitro experiments: gradient of the selection field 1.5 T m⁻¹, amplitude of all three drive fields 14 mT resulting in a sampling volume of 37  ×  37  ×  18.5 mm³. The dynamic MPI scan consisted of 35 000 frames acquired with the full temporal resolution of 21.5 ms. After 6000 frames the injections of 60 µl of tracer material were performed using the syringe pump.

Reconstruction and post-processing were performed offline. Reconstruction was performed again with the Kaczmarz algorithm. Frequency filtering was applied with a bandwidth of 0.08–1.25 MHz. In contrast to the in vitro examinations where 2000 averages over 43 s were used, in the in vivo setting we aimed at reaching the full temporal resolution of 21.53 ms. As a consequence reconstruction parameter had to be adjusted. We increased the SNR threshold from 1.4. to 1.7 and the number of iteration steps was set to 3. The regularization parameter was set to 5 * 10⁻² after systematically evaluating several parameters in the range of [10⁻⁴, 10²] on individual frames of the time series. Firstly, the entire dynamic sequence was reconstructed with a block averaging of 5, which improved the SNR and in turn the spatial resolution of the data. Since it also lowered the temporal resolution by a factor of 5, we additionally performed single frame reconstructions, in order to visualize the major vessels during the first pass after bolus injection and in order to study the blood pulsation. Image quality concerning the delineation of the MPI signal in the inferior vena cava and by the amount of background signal was judged by an observer on a subjective scale.

Finally, co-registration was performed manually using the same image registration framework as mentioned in the in vitro experiments. Alignment of both datasets was achieved by looking for the fiducial marker and anatomic landmarks of the inferior vena cava and the heart. Image fusion/windowing, ROI handling and curve fitting were performed with ImageJ (NIH, USA) extended by a plugin (qMapIt). ROI were placed in the heart, liver, and kidney. Signal-over-time curves were extracted and compared for LS-008 and Resovist. Further quantification was performed to estimate the perfusion fraction by calculating ratios of the signal in the ROIs of liver to blood and kidney to blood, respectively. As commonly performed (Nigrovic 1993), signal decay curves in blood were fitted to a bi-exponential model a₁ * exp(−m₁ * t)  +  a₂ * exp(−m₂ * t). The first 6 s after signal maximum were discarded to minimize the influences of the injection. Then the curves were normalized and a Levenberg–Marquardt algorithm was applied for the non-linear least square optimization. The amplitudes a₁ and a₂ reflect the amount of the injected tracer but were not further analyzed. More importantly, m₁ and m₂ are the rate constants for the distribution and clearance were discussed.

The two magnetic particle spectra of LS-008 and Resovist show decays in the amplitudes with higher frequencies (figure 1). In contrast to Resovist, LS-008 shows higher amplitudes and a longer decay which can be proven by the mean ratio of amplitudes of both spectra after a background subtraction of the noise floor. An averaged 3.4-fold amplification of LS-008 over Resovist was found.

Zoom In Zoom Out Reset image size

The SNR comparison of the system function data shows that LS-008 has higher SNR values over the full frequency spectrum than Resovist (figure 2). We found the same suitable SNR threshold for the Resovist and for LS-008 system function to be 1.4. With this threshold for Resovist 3909 frequency components in the x-channel were found, the highest at 1.217 MHz. For LS-008 we found 8216 frequency components at 1.244 MHz the highest. This means that LS-008 provides more than twice frequency components. The ratio of the filtered and summed SNR values given by ISNR_LS-008 (0 kHz, 1.25 MHz)/ISNR_Resovist (0 kHz, 1.25 MHz) was 3.7.

Zoom In Zoom Out Reset image size

After the system function analysis we calculated the maximum SNR (SNR_max) in the reconstructed phantom images to depict the impact of regularization. Plotting λ over SNR_max we generally found a decrease of SNR_max with decreasing concentration (figure 3(C)). Resovist shows a significant drop down at 3.2 mM which was less pronounced for LS-008 but also visible. For λ smaller than 10⁻² the situation worsens for the low concentration phantoms so we chose λ  =  10⁻² as a compromise for a direct image based comparison.

Zoom In Zoom Out Reset image size

Comparing the images one to one we can clearly depict a decrease in image quality with decreasing SPIO concentration for both tracers (figure 4(A)). The loss of image quality can be identified as a loss of detail, a loss of signal and an increased artifact level. First this occurs with the smallest hole containing the smallest amount of tracer material. Note that due to regularization noise visualizes as low resolution artifacts. When directly comparing both tracers, one can see that LS-008 performs clearly better than Resovist. With LS-008 four spots can be detected for concentrations higher 3.2 mM which is not possible with Resovist. Resovist depicts a significant noise increase at 1.1 mM while LS-008 still shows three delineated spots with less noise (figure 4(B)). At 0.4 mM the three spots cannot anymore be surely depicted as noise becomes more pronounced.

Zoom In Zoom Out Reset image size

We analyzed the applicability of the tracers for MPI angiography. To show the reproducibility of the MPI measurements the inflow of all six examination are shown in figure 5. Both tracers could visualize the propagation of the bolus through the inferior vena cava in all six animals. Concerning the artifact level there was a clear visible quality difference (figures 5 and 6). Compared to Resovist, LS-008 clearly showed fewer temporally fluctuating artifacts probably due to the higher SNR. There were other pronounced differences in tracer performance. In frames showing the bolus leaving the heart, LS-008 clearly distinguished the aorta form the caval vein (figure 6(C)); furthermore, a vessel passing the liver and a cerebral vessel structure leading cranial was also observed (figure 6(D)). Both anatomical details were missing in images acquired with Resovist. We also examined the signal change in the caval vein triggered by periodic cardiac and respiratory motion. Again, LS-008 offers more details than Resovist (figure 6(E)) as the signal modulation caused by breathing can be clearly depicted. This modulation was visible during the whole dynamic scan time and might me exploited in future to track vital signs. The non-periodic temporal variations were generally higher for Resovist.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

We examined the organ perfusion and SPIO distribution. Using LS-008 the visualization of the liver and kidneys as well perfused organs was successful (figures 7(A) and (B)). With Resovist kidneys could not be surely identified. A clear difference in tracer distribution between LS-008 and Resovist was depicted after 8 min (figures 7(C) and (D)). Resovist, formally developed for liver diagnostics, showed after the arrival of the bolus a decaying signal in blood while at the same time a signal increase in liver (figure 8). In contrast, the signal decay with LS-008 in blood was slower. Furthermore, after an initial perfusion the signal levels remained constant in liver and kidney. Based on the ROI analysis of the LS-008 data we calculated the perfusion fraction, defining the part of a voxel containing vessels and capillaries, being approximately 30% for liver and 22% for kidneys (figure 8).

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Modeling a bi-exponential function to the decaying signal of the intracardial blood we depicted rate constants (mean value  ±  standard deviation). For LS-008 we found m₁  =  4.30  ±  1.54 min⁻¹, m₂  =  0.0079  ±  0.0067 min⁻¹ and for Resovist m₁  =  1.10  ±  0.49 min⁻¹, m₂  =  0.0519  ±  0.0089 min⁻¹. m₂ reflects the slower phase which is the blood clearance; m₁ the fast distribution after injection. For LS-008 the blood half-life of the slow phase is 88 min and for Resovist, respectively, 13 min. In figure 9 the signal over time curves generated from the rate constants are shown. Here the 50% signal drop for Resovist is reached around 5 min. For LS-008 we extrapolated a 50% signal drop to be at 77 min.

Zoom In Zoom Out Reset image size