Brain Iron Detected by SWI High Pass Filtered Phase Calibrated with Synchrotron X-Ray Fluorescence

Abstract

Purpose

To test the ability of susceptibility weighted images (SWI) and high pass filtered phase images to localize and quantify brain iron.

Materials and Methods

Magnetic resonance (MR) images of human cadaver brain hemispheres were collected using a gradient echo based SWI sequence at 1.5T. For X-ray fluorescence (XRF) mapping, each brain was cut to obtain slices that reasonably matched the MR images and iron was mapped at the iron K-edge at 50 or 100 μm resolution. Iron was quantified using XRF calibration foils. Phase and iron XRF were averaged within anatomic regions of one slice, chosen for its range of iron concentrations and nearly perfect anatomic correspondence. X-ray absorption spectroscopy (XAS) was used to determine if the chemical form of iron was different in regions with poorer correspondence between iron and phase.

Results

Iron XRF maps, SWI, and high pass filtered phase data in nine brain slices from five subjects were visually very similar, particularly in high iron regions. The chemical form of iron could not explain poor matches. The correlation between the concentration of iron and phase in the cadaver brain was estimated as c_Fe [μg/g tissue] = 850Δφ + 110.

Conclusion

The phase shift Δφ was found to vary linearly with iron concentration with the best correspondence found in regions with high iron content.

There is a pressing need to establish better techniques to quantify iron in vivo in the most common neurodegenerative disorders such as Alzheimer’s disease, Parkinson’s disease, and multiple sclerosis where it is thought that iron plays a key role in the disease pathology (1,2).

Several methods have been developed to non-invasively image iron distribution using magnetic resonance imaging (MRI). T2* (3), T2 (4), and T2′ (5) have shown some success in quantifying iron but there is wide dispersion in mapping iron with these techniques. Magnetic field correlation imaging is also sensitive to iron content and has been shown to correlate well with average nonheme iron values (6–9) as reported by Hallgren and Sourander (10,11). Field dependent relaxation rate increase (FDRI) is correlated with iron concentration, but requires imaging the same subject at different field strengths on different scanners making it difficult to acquire the data practically (12). Over the last few years, a strong case has been made that susceptibility weighted imaging (SWI) (13,14) is sensitive to iron content and more specifically that SWI high pass filtered phase images can be used to measure iron content (15,16).

SWI was originally developed to track blood oxygen saturation and image veins (13) and is now widely used clinically for a variety of neurovascular-related diseases and to infer iron content (14,17–20). It has been used successfully to image vascular malformations in Sturge-Weber syndrome (21) and diseases with potential iron involvement such as multiple sclerosis (22). The brain accumulates iron with age (10) and several studies have employed SWI to examine baseline morphology (20) and changes associated with aging (23,24). Although phase has been used to measure putative iron content and has been correlated to published iron concentrations (20), a direct relationship between iron concentration and phase of the MR signal in the same brain has not been demonstrated and the relationship between phase and iron has not been calibrated.

SWI utilizes a T2*-weighted gradient echo sequence and further enhances magnetic susceptibility differences by incorporating acquired phase information (15). This method differs from other techniques that use only the magnitude of the MR signal and ignore its relative phase. While phase information is influenced by differences in magnetic susceptibility, the phase images themselves are difficult to interpret. Since recorded angles are limited to values between −π and π, areas of high susceptibility will result in aliasing. Interfaces between areas with large differences in magnetic susceptibility, such as between air and tissue, also result in large background phase variations that can obscure smaller local phase changes inside the tissue itself. The local tissue susceptibility is made clinically accessible by homodyne filtering to reduce low spatial frequency variations in the phase image. The size of filter used effects the resulting phase (20).

The filtered phase information is then used to generate a mask that is multiplied onto the acquired magnitude image to highlight areas of high magnetic susceptibility in the context of local anatomy (20). The image produced by this process will be referred to as an “SWI image” in this paper. Although a minimum intensity projection (mIP) SWI image combining three or four slices is often used clinically to enhance the visibility of iron-rich structures (especially blood vessels), we have chosen not to display multi-slice images because it was more important to find a close match to the XRF slice from among the MR image stack.

X-ray fluorescence (XRF) imaging is a well-established, element-specific technique for quantitative mapping of multiple metals that has recently been modified to map large brain slices (25–27). In this study, we compared and found good correspondence between iron distribution as determined by XRF and both SWI and phase. We quantified iron using an XRF iron standard and compared it to the original magnitude and filtered phase data from well matched MR slices from the same brain to determine the relationship between phase and brain iron concentration.

MATERIALS AND METHODS

XRF Imaging and Spectroscopy

Rapid-scanning XRF imaging (RS-XRF) and X-ray absorption spectroscopy (XAS) are well established, element-specific, quantitative techniques used to determine the location and chemical form of elements in a broad range of samples including whole tissues.

For imaging, the incident X-ray beam energy is fixed so as to excite all elements with K-shell absorption edges below the incident energy. By carefully windowing on the XRF emission lines, the location of multiple elements can be simultaneously mapped. XRF is vastly superior to histochemical staining for metals because it relies upon the physics of the atom rather than a chemical reaction and therefore total iron can be accurately mapped regardless of chemical form. Moreover, because the X-ray beam is on the sample for milliseconds, radiation damage is low and the sample can subsequently be prepared for other analytic techniques such as histology or MRI. The related technique of XRF microprobe, using a smaller beam and higher count times has been used to detect metals in individual cells (28) and small regions of tissue (29–31).

The chemical form of iron was determined in cubes of fixed brain tissue cut from brain slices that had been previously imaged. XAS spectra were collected at 10K using a liquid helium-cooled cryostat. The incident beam energy was scanned across the iron K absorption edge and fluorescence was detected with a 30 element Ge detector (Canberra). Since fixation alters iron’s chemical form, the nature of the ligands in different brain regions was compared by overlaying the K-edge spectra.

SAMPLE PREPARATION

Five formalin-fixed human cadaveric brain hemispheres were obtained from the Saskatoon Movement Disorders Clinic brain repository. The clinical diagnoses were confirmed by post-mortem pathology. Hemispheres were placed in warm 10% gelatine under vacuum to remove adherent air bubbles and then kept chilled.

MRI

MR images were collected using a gradient echo based SWI sequence (α = 15°, TE = 20 msec, TR = 28 msec) on a 1.5T Siemens Magnetom Symphony system. The data was acquired in transverse and coronal orientations using a circularly polarized head coil. Images were acquired with 0.5 mm resolution in the phase and readout directions, and 2 mm in the slice direction with a bandwidth of 120 Hz/pixel, a field-of-view of 256 mm × 256 mm and displayed on a 512 × 512 matrix. MR phase images were post-processed using high pass filtering with a central 32 × 32 matrix using SPIN software (Signal Processing in NMR, Detroit, MI, USA). Zero phase (2048 phase units) was confirmed to be gelatin based upon three measurements collected from gelatin from each of two different brain slices, that yielded an average of 2051 ± 12 phase units.

RS XRF Sample Selection

Each brain was manually cut transversely and then in the coronal orientation so as to closely match an MR slice. A 1-mm thick slice was then taken from each cut face using an industrial meat slicer. Each slice was sprayed with 10% buffered formalin and heat-sealed under 3-μm thick Mylar (Goodfellow Metals) for XRF mapping.

XRF Metal Mapping

RS-XRF imaging was done on beamline 10-2 at the Stanford Synchrotron Radiation Lightsource which is equipped with a 30 pole 1.45 T wiggler source. The incident beam was set at 13 keV and was vertically collimated onto a Si (111) monochromator. The incident beam intensity was monitored using a nitrogen-filled ion chamber upstream of a tantalum aperture. This aperture, producing a 50 μm × 50 μm spot on the sample, was used for initial mapping (Figs. 1 and 2) while a 100 μm aperture was used to improve signal to noise for the quantification study (Fig. 3). The samples were mounted on a set of motorized stages (Newport Corp.) oriented at 45° to the incident beam to minimize scatter, and raster scanned in the beam with a dwell time of approximately 12.5 msec/point. A single element silicon drift detector (Vortex) was placed at 90° to the incident beam and manually aligned with the intersection point of the beam and the sample. Energy windows were centered to collect fluorescence counts from the emission lines of Fe Kα₁ and Kα₂, all other biologically interesting elements, scatter and total incoming counts. Fe emission lines were averaged at a mean Kα energy of (2E_Kα1 + E_Kα2)/3.

Figure 1.

Brain slices showing good correspondence between SWI and XRF (a, b, e, f) and filtered phase and XRF (c, d, g, h). Progressive supranuclear palsy transverse (a, b) and coronal (c, d); Parkinson’s disease transverse (e, f) and coronal (g, h); c, caudate; cp, choroids plexus; gp, globus pallidus; p, putamen; sn, substantia nigra; rn, red nucleus; t, thalamus; * white matter. Scale bar = 10 mm.

Figure 2.

Brain slices showing modest correspondence between SWI and XRF (a, b, e, f) and filtered phase and XRF (c, d, g, h). Progressive muscular atrophy transverse (a, b) and coronal (c, d); Cerebrocranial trauma transverse (e, f), coronal (g, h); c, caudate; p, putamen; sn, substantia nigra; ic, internal capsule; t, thalamus; * white matter. Scale bar = 10 mm.

Figure 3.

Iron Quantification on Alzheimer’s disease coronal section. a: SWI. b: XRF iron map. c: High pass filtered phase.

XRF images were prepared for display using Interactive Data Language (IDL) as described previously (27) except that the highest and lowest values were retained for quantification purposes.

Iron Quantification

Metal concentrations were determined using an XRF iron foil standard (Micromatters Inc.). In thick samples, such as the brain slices used here, the detector registers fluorescence counts when the incident X-ray beam meets the sample, penetrates to some depth, interacts with absorber atoms and produces fluorescence photons by the photoelectric effect. The XRF escapes in all directions with some along a path toward the detector. Some absorption of the incident X-ray beam will occur as it travels into the sample before arriving at the fluorescing material. There is also an attenuation of the fluorescence photons as they escape to the detector. With a thick sample, we must therefore take into account the fact that the fluorescence counts from the fluorescing material, in this case, iron, will decrease as the iron is more deeply embedded into the sample, to the point where either no incident X-rays strike an iron molecule or no fluorescence photons escape the sample. For iron, this occurs in only a few hundred μm (1/e attenuation occurs at 310 μm in brain tissue).

The iron calibration foil is thin enough that these thickness effects do not apply. Neglecting attenuation by air, there is no absorption of the incident beam before arriving at the sample, and no attenuation of the fluorescence as it escapes toward the detector. Taking these factors into account, we have arrived at the following formula to estimate the concentration of a given metal in a thick sample based on fluorescence counts from a calibration foil:

$$N_{F_{\mathit{sample}}}=\frac{n_{F_{\mathit{sample}}}}{n_{F_{\mathit{foil}}}}\frac{{\mathrm{\sum }}_{F_{\mathit{foil}}}{M}_u}{N_A{\mathrm{\nu }}_{u}sin\mathrm{\theta }}\frac{\mathrm{\mu }}{{\mathrm{\rho }}_T}(E_0+E_{K}tan\mathrm{\theta }),$$ [1]

where N_Fsample is the number density of fluorescing atoms in the sample (i.e., number of iron atoms), n_Fsample is the fluorescence count from the sample (measured by the detector), n_Ffoil is the fluorescence count from the foil standard, Σ_Ffoil is the number density per cross-sectional area of fluorescing atoms in the foil, M_u is the molar mass constant = 1 g/mol, N_A is Avogadro’s constant, ν_u is u/ρ where u is the atomic mass unit (= 1.66 × 10⁻²⁴ g) and ρ is sample density, $\frac{\mathrm{\mu }}{{\mathrm{\rho }}_T}$ is the mass attenuation coefficient due to photoelectric effect, E₀ is the energy of the incident beam, E_K is the fluorescence energy, θ is the incident angle of the beam.

At synchrotron facilities, beam flux varies, depending on whether or not the electron beam in the storage ring has recently been “topped up”. For our experimental setup with an incident x-ray beam angle of 45°, incident energy of 13 keV, calibration foil area density of 46 μg/cm², and a brain tissue density of 1.04 g/mL, the average number density would be

$$N_{F_{\mathit{sample}}}=1.7\times {10}^{19}\left(\frac{n_{F_{\mathit{sample}}}}{n_{F_{\mathit{foil}}}}\right),$$ [2]

or iron concentration in μg/g,

$$c_{Fe}[\mathrm{\mu }\mathrm{g}/\mathrm{g}\text{tissue}]=1490\left(\frac{n_{F_{\mathit{sample}}}}{n_{F_{\mathit{foil}}}}\right).$$ [3]

Quantification done using the standard XRF foil and iron solutions in 2-mm thick Lucite sample cells with scatter similar to brain tissue were within 16% of each other.

The relationship between number density of fluorescing atoms and phase was calculated from 17 anatomic regions from a subject with Alzheimer’s disease that were defined in the raw XRF images and corresponding SWI and high pass filtered phase images. XRF counts and corresponding iron concentration were plotted against phase in corresponding areas (Fig. 4). The relationship between iron and phase was analyzed by least squares linear regression using GraphPad Prism 5. The populations were non-Gaussian and so associations between XRF and phase were also evaluated by the Spearman-rank correlation coefficient.

Figure 4.

Relationship between phase and fluorescence. Linear regression coefficients: r² = 0.81, P < 0.0001.

RESULTS

Four brains were sliced in transverse and coronal orientation to match SWI and high pass filtered phase images respectively taken of the intact brain hemisphere. An additional coronal section was mapped for quantification purposes. These five cases are considered below. When XRF, SWI, and filtered phase images were visually compared there was excellent correspondence in some images (Figs. 1 and 3) and modest correspondence in others (Fig. 2).

Progressive Supranuclear Palsy

In the transverse orientation (Fig. 1a and b) excellent correspondence is seen in the substantia nigra and the iron-rich U-fibers at the interface between the relatively iron-poor outer gray matter and adjacent cortical white matter. The choroid plexus is resolved in both XRF and SWI but is very hypointense in the SWI image (Fig. 1b, black arrowheads). In the coronal section (Fig. 1c and d) the basal ganglia have been magnified to highlight the excellent correspondence between XRF and phase images in both the location and concentration of iron in the putamen, caudate, and globus pallidus (Fig. 1 c, d, arrows).

Parkinson’s Disease

Good correspondence is seen between the SWI image (Fig. 1e) and the XRF slice (Fig. 1f). The iron-rich U-fibers and substantia nigra (arrowhead) are visible in both the XRF and SWI image but the iron-rich red nucleus and thalamus (arrows), clearly seen in the XRF image, are not well visualized by SWI. In the coronal slice (Fig. 1g and h), the pattern of hyperintensity in the phase image of the caudate nucleus, putamen, and globus pallidus (arrows) corresponds very well to the XRF iron map. Whereas the XRF iron image shows no increase in iron in the lateral white matter relative to the internal capsule, the internal capsule appears brighter in the phase image (possibly due to filtering effects (32)).

Progressive Muscular Atrophy

The transverse SWI slice is not perfectly matched to the XRF slice (Fig. 2a and b) and this may, at least in part, account for the poor correspondence but as in Fig. 1, the choroid plexus is very hypointense in the SWI image (Fig. 2a). There is excellent correspondence between the phase and XRF images (Fig. 2c and d) since the imaging planes are almost identical as shown by the sulci of the frontal lobe. A hyperintense border along the lateral and inferior aspects of the basal ganglia and hypointensities centrally in the phase image correspond nearly perfectly to a peripheral region high in iron and a central region lower in iron in the XRF map (Fig. 2c and d). There is also a good correlation between iron and phase in the body of the caudate (Fig. 2c and d), but adjacent white matter (Fig 2 c, d, asterisk) shows less iron than the phase image indicates.

Craniocerebral Trauma

While the shape of the substantia nigra in the XRF and SWI images are well matched in the transverse section (Fig. 2e and f), the intensity of the signal arising from the central white matter is dramatically different (Fig. 2e and f, asterisk). The coronal slice shows modest correspondence between XRF and phase in the location of iron in the basal ganglia (Fig. 2g and h) but poor quantitative correspondence.

Alzheimer’s Disease

An additional coronal slice was mapped at 100 μm resolution, using a longer 30 msec count time along with an XRF iron standard. This slice was chosen for quantification because it showed a large range of iron concentrations by XRF and the anatomic match between MR and XRF slices was nearly perfect. Both SWI (Fig. 3a) and high pass filtered phase (Fig. 3c) are shown in addition to XRF (Fig. 3b). Overall the correspondence between XRF and SWI is very good. However, the phase arising from the internal capsule is lower than that from the central white matter whereas XRF shows no difference in iron content between the internal capsule and central white matter. Again this may be due to filtering effects of local phase enhancement coming from geometry-dependent magnetic field effects.

Relationship Between Iron and Phase

Regions of white matter and gray matter were defined in the XRF and high pass phase images and the average fluorescence counts were plotted against phase units (Fig. 4). The Spearman correlation coefficient (rho) was 0.786 and P = 0.0002.

A relationship was found between fluorescence counts and phase in radians (Δφ):

$$\mathrm{\Delta }\mathrm{\phi }=0.0025n_{F_{\mathit{sample}}}-\mathrm{0.13.}$$ [4]

Since our average foil count was 700, Eq. [3] gives a calibrated relationship between iron concentration and fluorescence:

$$c_{Fe}[\mathrm{\mu }\mathrm{g}/\mathrm{g}\text{tissue}]=2.13n_{F_{\mathit{sample}}}.$$ [5]

Fluorescence and phase with calculated iron concentrations using the average foil count are shown in Table 1.

Table 1.

Iron concentration for selected regions of the Alzheimer’s brain calculated from fluorescence and phase measurements

Anatomic region	XRF (counts)	Phase (radians)	Iron μg/g wet weight (average foil)
Periventricular white matter	21.8	0.0002	46
Subcortical white matter, precentral gyrus	39.6	−0.002	84
Superior temporal gyrus	41.2	−0.003	88
Insular cortex	43.0	−0.055	92
Middle frontal gyrus	43.0	−0.005	92
Cortex, precentral gyrus	47.3	−0.020	101
Internal capsule	52.0	−0.175	111
External capsule and claustrum	54.8	0.009	117
Middle temporal gyrus	57.5	−0.023	122
Ventricular wall, superior to caudate nucleus	67.5	0.035	144
Superior caudate nucleus	91.5	0.160	195
Superior globus pallidus	121.7	0.166	259
Superior putamen	122.5	0.114	261
Inferior caudate nucleus	123.1	0.189	262
Anterior commissure	144.0	0.207	306
Inferior globus pallidus	157.6	0.307	335
Inferior putamen	163.5	0.268	348

By combining Eqs. [4] and [5], iron concentration as a function of phase may be written as

$$c_{Fe}[\mathrm{\mu }\mathrm{g}/\mathrm{g}\text{tissue}]=850\mathrm{\Delta }\mathrm{\phi }+110.$$ [6]

Since SWI imaging was performed at 1.5T with an echo time of 20 msec, a more universal form of this equation would be:

$$c_{Fe}[\mathrm{\mu }\mathrm{g}/\mathrm{g}\text{tissue}]=850\mathrm{\Delta }\mathrm{\phi }\left(\frac{1.5}{B_0[T]}\right)\left(\frac{20}{TE[\text{ms}]}\right)+110.$$ [7]

As shown in Table 1, using the average XRF counts from the foil (700), the highest iron concentration in the Alzheimer’s brain was 335 μg Fe/g and the lowest was 46 μg Fe/g. These values are slightly higher than normal brain iron concentrations as determined by other methods and may be related to disease or advanced age.

Iron Chemistry

There was poor correspondence between XRF and SWI in some regions and so we sought to determine the cause. Since XRF detects all iron and MRI detects only paramagnetic iron, we hypothesized that a change in the ratio of paramagnetic ferritin to non-paramagnetic iron species might reduce the MR signal in some regions or disease states. To determine if the iron K-edge spectra differed between brain regions showing excellent or poor correlation between XRF and SWI, blocks of tissue were excised from the brains shown in Figs. 1 and 2 for near-edge analysis of K-edge spectra (Fig. 5). The ferritin and heme spectra are quite distinct (Fig. 5d) and therefore changes in their proportions within brain tissue would change the shape of the K-edge spectrum. We found that white matter and gray matter, the two major tissue types in brain (Fig. 5a), regions with poor matches (cortical gray matter and red nucleus) (Fig. 5b) and regions with perfect matches (putamen and globus pallidus pars externa) (Fig. 5c) could be overlaid with each other indicating that changes in MR signal are not due to regional differences in iron species.

Figure 5.

The chemical form of iron in different brain regions is very similar. Iron K-edge spectra are ‘fingerprints’ of chemical form. Spectra were collected from cubes of brain tissue from the slices shown in Figs. 1 and 2. a: overlay of spectra from major tissue types, white matter and gray matter; (b) overlay of spectra from brain regions with poor correspondence between MR and XRF, cortical gray matter and red nucleus; (c) overlay of spectra from brain regions with excellent correspondence between MR and XRF, putamen and globus pallidus pars externa; (d) overlay of model iron compounds of species abundant in the brain, ferritin and hemoglobin.

DISCUSSION

We have shown that within the same brain slice, regions of hyperintensity in high pass filtered phase MR images and SWI correlate reasonably well with regions having high iron Kα fluorescence. Because iron Kα fluorescence is linear with iron concentration, we were able to determine the relationship between iron concentration and phase response. Although it is true that objects with uniform susceptibility differences will be quantifiable by susceptibility mapping, we have adopted the phase approach because phase shifts could be due to chemical shift in which case susceptibility mapping will fail. However, for macroscopic susceptibility differences, the phase inside any object (and outside) will depend on its geometry and the orientation of that geometry to the main field. This can lead to changing fields depending on the orientation of the brain to the main magnetic field. This variation could account for some of the dispersion seen in our results.

Since the registration between images was only approximate in many cases, we calculated the relationship between iron concentration and phase by selecting specific anatomic regions and averaging pixel intensities within those regions (Table 1) in the brain slice with the best anatomic correspondence. The relationship between phase and iron given in Eq. [7] predicts somewhat more iron per unit phase than was predicted in Haacke et al (16). There, the expression

$$c_{Fe}[\mathrm{\mu }\mathrm{g}/\mathrm{g}\text{tissue}]=217\mathrm{\Delta }{\mathrm{\phi }}_{\mathit{CSF}},$$ [8]

was assumed with Δφ_CSF being the phase difference between the tissue of interest and cerebrospinal fluid (at 1.5T and 40 msec, CSF was assumed to have an offset of −0.112 radians from the origin, which in most cases of brain imaging is represented by the white matter signal). In our study the brains were embedded in gelatine prior to MR. Following the protocol described in Haacke et al (16) we determined that the Siemens MRI was setting the signal from gelatin to have a phase of zero radians (2048 phase units). As gelatin has very little MR-detectable iron, we should not need to adjust the y-intercept of Eq. [4].

Now, taking into account the field strength, the echo time and the offset Eq. [8] becomes

$$c_{Fe}[\mathrm{\mu }\mathrm{g}/\mathrm{g}\text{tissue}]=434\mathrm{\Delta }{\mathrm{\phi }}_{\mathit{CSF}}\left(\frac{1.5}{B_0[T]}\right)\left(\frac{20}{TE[\text{ms}]}\right).$$ [9]

Comparing Eq. [7] to Eq. [9] we see that the slope for iron is two times higher than estimated in Ref. 11. This implies that it takes more iron (by a factor of 2 higher than originally anticipated in Ref. 11) to generate a given phase shift.

Iron concentration is not the only factor affecting phase. In brains with high iron content, iron will contribute proportionately more to the phase than other factors thus increasing the slope. While the brain quantified in this study had above normal iron content, iron was in agreement with previous analyses of Alzheimer’s disease brain for non-heme iron (11) and total iron (33).

White matter has low iron content relative to other parts of the brain and so the contribution of iron to total phase is necessarily lower and other factors such as water content contribute more. We found that the relationship between iron and phase was weaker for white matter. Gelman et al (5) show that the error in R2′ in white matter is more than double that in gray matter and white matter values are skewed above the linear fit. In our experiment, white matter values are more scattered. This may result from changes in water content due to formalin fixation (34) or the orientation of white matter fibres as described by He and Yablonskij (35).

We observed regions in which brain iron as measured using XRF did not match SWI either visually or quantitatively. Several factors could lead to such discrepancies. These, discussed below, include poor anatomic correspondence, voxel size, iron chemistry (paramagnetic vs. non-paramagnetic), slice-to-slice misregistration, high pass filter effects on larger structures and finally the presence of other materials such as calcium that might affect the phase.

Mismatched section planes can result in poor correspondence between MR and XRF images and this is demonstrated by comparing Figs. 1 and 2. We have recently attempted to collect XRF maps prior to high resolution MR on the slice. This strategy has its own technical challenges but may prove to be more successful (data not shown).

Phase behavior does not always match iron concentration implying that the relationship between phase and iron is not linear over the entire range of concentrations. We would expect a better correspondence between XRF and MRI if the size of the voxel were the same. In the current study the voxel size of XRF was about 100 (horizontal) × 100 (vertical) × 310 (deep) μm while the MR voxel was 500 μm resolution in the phase and readout directions, and 2000 μm in the slice direction. The much bigger MR voxel (over 100×) would average high-iron areas with adjacent lower-iron areas. We attempted to overcome the voxel size mismatch by averaging pixels/voxels within defined anatomic areas but the data does not enable us to overcome the difference in depth profile between the two techniques. Although difficult to quantify in tissue as heterogeneous as brain, the high resolution and shallow depth profile of XRF will necessarily yield a higher value for iron in thin regions with uniformly high iron content than will MR.

The chemical form of iron could also play a role since XRF detects all iron in all oxidation states and chemical forms whereas MR only detects iron that is in a magnetic state. The living brain has abundant paramagnetic deoxyheme that is not found in the preserved postmortem brain and therefore in our experiments there would be no contribution from deoxyheme iron to the phase. We tested the hypothesis that weakly magnetic or non-magnetic iron such as heme, transferrin or iron-sulfur cluster proteins might be abundant in regions with poor correspondence between XRF and SWI. Spectra from all brain regions could be overlaid indicating that the mix of iron species is similar from region to region.

The highly-ordered six-line ferrihydrite mineralized by ferritin is known to be superparamagnetic, but as the ferritin protein shell degrades during the formation of hemosiderin (36) some of the iron from the ferritin cores may become free and precipitate as amorphous ferrihydrite. The magnetic properties of amorphous ferrihydrite are complex (37) and need to be empirically determined in the brain. Our spectra could not distinguish between six-line and amorphous ferrihydrites and XAS does not have the resolution to detect the tiny amounts of magnetite that are known to be present in normal (38) and AD brain (39–41) and that could have an effect on phase.

The chemical form of iron needs to be considered when comparing XRF with histological methods of detecting iron. While most of the iron in brain is found in ferritin or hemosiderin which are detectable with Perls stain, this preponderance of ferric iron may not be true for all tissues or disease states. In contrast to Perls stain, XRF permits us to generalize our calibration beyond the brain.

The high pass filtering method used to remove the background phase shifts caused by the air-gelatin interface may suppress the phase in objects with significant low spatial frequency components (those on the order of 1 cm or more in dimension). A new method developed by Neelavalli et al (42) computes and removes phase shifts caused by the susceptibility gradient at the air interface. This method could reduce or eliminate the confounding influence of most of the high pass filtering effects.

We must also consider the effect of elements other than iron on the MR signal. In particular the signal from calcium may lead to an underestimation of iron in calcium-rich brain areas. Iron compounds tend to be paramagnetic and thus generate weak local fields that add to the strength of the externally applied magnetic field. Calcium compounds, which tend to be diamagnetic, subtract from the strength of the external field. The collocation of calcium with iron could reduce the phase shift caused by iron, thereby underestimating the actual iron concentration (43). Further work is required to fully elucidate the effect of brain calcium on the detection of iron by MRI.

In conclusion, by comparing SWI and filtered phase with XRF iron mapping, we have shown that SWI provides a reasonably reliable approach to localize iron and compare its concentration. The phase shift Δφ was found to vary linearly with iron concentration with the best correspondence found in regions with high iron content.