High resolution anatomical and quantitative MRI of the entire human occipital lobe ex vivo at 9.4T.

Several magnetic resonance imaging (MRI) contrasts are sensitive to myelin content in gray matter in vivo which has ignited ambitions of MRI-based in vivo cortical histology. Ultra-high field (UHF) MRI, at fields of 7T and beyond, is crucial to provide the resolution and contrast needed to sample contrasts over the depth of the cortex and get closer to layer resolved imaging. Ex vivo MRI of human post mortem samples is an important stepping stone to investigate MRI contrast in the cortex, validate it against histology techniques applied in situ to the same tissue, and investigate the resolutions needed to translate ex vivo findings to in vivo UHF MRI. Here, we investigate key technology to extend such UHF studies to large human brain samples while maintaining high resolution, which allows investigation of the layered architecture of several cortical areas over their entire 3D extent and their complete borders where architecture changes. A 16 channel cylindrical phased array radiofrequency (RF) receive coil was constructed to image a large post mortem occipital lobe sample (~80×80×80mm3) in a wide-bore 9.4T human scanner with the aim of achieving high-resolution anatomical and quantitative MR images. Compared with a human head coil at 9.4T, the maximum Signal-to-Noise ratio (SNR) was increased by a factor of about five in the peripheral cortex. Although the transmit profile with a circularly polarized transmit mode at 9.4T is relatively inhomogeneous over the large sample, this challenge was successfully resolved with parallel transmit using the kT-points method. Using this setup, we achieved 60μm anatomical images for the entire occipital lobe showing increased spatial definition of cortical details compared to lower resolutions. In addition, we were able to achieve sufficient control over SNR, B0 and B1 homogeneity and multi-contrast sampling to perform quantitative T2* mapping over the same volume at 200μm. Markov Chain Monte Carlo sampling provided maximum posterior estimates of quantitative T2* and their uncertainty, allowing delineation of the stria of Gennari over the entire length and width of the calcarine sulcus. We discuss how custom RF receive coil arrays built to specific large post mortem sample sizes can provide a platform for UHF cortical layer-specific quantitative MRI over large fields of view.

Introduction

With increases in main field strengths, from 3 T to 7 T and beyond, and increases in the number of channels in phased-array RF coils, it has become increasingly clear that RF coils specifically built to subject or sample size and shape can offer very large advantages (de Zwart et al., 2002, Wiggins et al., 2006, Keil et al., 2011, Oezerdem et al., 2016). Although this has specially been true for human in-vivo 7 T coil designs, this approach is less developed in the field of ex-vivo tissue MR. Investigations of human ex-vivo brain samples in particular have therefore been limited to either tissue samples that fit pre-clinical MR setups, or the contrast and Signal-to-Noise ratio that can be achieved with a non-specialized setup, such as an in-vivo head coil.

Recently a number of investigations have shown that several magnetic resonance imaging (MRI) contrasts (such as T1, T2, T2* and phase) show sensitivity to iron and, especially, myelin content in gray matter in vivo (Duyn et al., 2007, Glasser and Van Essen, 2011, Barazany and Assaf, 2012, Dick et al., 2012, Sereno et al., 2013, De Martino et al., 2014, Lutti et al., 2014, Stuber et al., 2014) which has ignited ambitions of MRI-based in vivo histology (Dick et al., 2012, Deistung et al., 2013, Sereno et al., 2013, Truong et al., 2014, Turner and Geyer, 2014). In some of these investigations quantitative MRI, providing quantitative contrasts in physical units rather than T1 or T2* weighted acquisitions, has been shown to have clear advantages for cortical MRI based histology (Dick et al., 2012, Deistung et al., 2013, Dinse et al., 2013). However, so far in vivo resolution, even at UHF, has not been sufficient to resolve individual cortical layers. This is why many studies use ex vivo MRI of human post mortem samples as a stepping stone to investigate MRI contrast in the cortex, validate it against histology techniques applied in situ to the same tissue, and investigate the resolutions needed to translate ex vivo findings to in vivo UHF MRI (Geyer et al., 2011, Augustinack et al., 2013a, Augustinack et al., 2013b, Augustinack et al., 2014, Turner and Geyer, 2014, Waehnert et al., 2014, Modo et al., 2016).

Ex vivo imaging involves scanning post-mortem tissue over long scan sessions (several hours to a couple of days) at UHF which can yield ultra-high resolution datasets far below the millimetre scale while maintaining a sufficiently high Signal-to-Noise ratio (SNR) for the acquired images. Ex vivo UHF MRI to examine human brain tissue samples has been predominantly performed on preclinical or animal MRI systems. (e.g. Glover et al. 1994; Fatterpekar et al. 2002; Roebroeck et al. 2008; Utz and Monazami 2009; Nabuurs et al. 2011; Seehaus et al. 2013; Aggarwal et al. 2015; Calabrese et al. 2015). This takes advantage of the high field strength & increased gradient performance of preclinical systems but is often limited to small human tissue samples (smaller than about 20×20×20 mm3), restricting size of the brain region examined. Alternatively, relatively small tissue samples have been investigated on a human 7 T MRI system either using a commercial head RF coil or a small custom built RF coil (e.g. Augustinack et al., 2013a, Augustinack et al., 2013b; Augustinack et al. 2014; Weiss et al. 2015). Extending such UHF studies to large human brain samples, while maintaining high resolution, would be an important step because it would allow investigations of the architecture of several cortical areas over their entire extent and their complete borders where architecture changes, which is very relevant to the translation to the in vivo situation. This requires sample-size specific RF-coils used in a large bore system (Augustinack et al., 2005, Augustinack et al., 2013a, Augustinack et al., 2013b) which would be difficult in a preclinical system due to space constraints.

Therefore, in this study, we focus on a custom built UHF RF-coil as a key technology for high resolution ex vivo anatomical and quantitative MRI. We designed and characterised a cylindrical phased-array receive (Rx) coil for use in a large-bore 9.4 T system and compared its performance with that of a standard whole head array coil at 9.4 T. The two specific aims in enabling cortical laminar discrimination over large FoVs are the following. First, we aim to achieve high resolution anatomical images (<100 μm) for a large (~80×80×80 mm3) human occipital lobe sample. Second, we aim to achieve sufficient control over SNR, B and B homogeneity and multi-contrast sampling to perform quantitative T2* imaging over the same volume at 200 μm. We use the primary visual cortex along the entire calcarine sulcus as a showcase to illustrate the capacity to delineate architectural details over a large surface of cortex. We also highlight how the interaction of RF coil hardware, B homogenisation and data analysis must be tuned to achieve optimal results.

Methods

Mechanical RF-coil and sample container construction

The coil was constructed on a hollow cylindrical transparent polycarbonate (PC) former and consists of 16 receive coils built in a phased-array layout on the outer surface of the hollow cylinder (Fig. 1A). The receive array former has an external diameter of 90 mm and a wall thickness of 3 mm, allowing an inner diameter of 84 mm to accommodate ex vivo samples. The receive coils were constructed using 1 mm thick copper wires coated with a thin layer of polyimide for insulation – with each individual coil being 42 mm in diameter – and laid out in a 8×2 grid of 16 coils, providing a cylindrical region of interest for imaging purposes. A critical overlap (~0.20 to 0.25 times the loop diameter of each coil) was maintained between all coils to minimize mutual inductance and to ensure mutual decoupling (Roemer et al. 1990). A layer of Kapton was taped over each coil such that the overlapping areas of each coil pair would be insulated from each other, while two distributed capacitors for each receive loop were divided symmetrically along the coil loop. A detailed schematic of the individual receive coil element is shown in Fig. 1D.

Fig. 1

(A) The 16 channel ex vivo sample receive RF coil with preamplifiers and cable trap assemblies (B). The receive coil integrated with a 16 channel transmit coil (C) The 3D-printed sample container with the occipital lobe sample immersed in Fluorinert (D). Circuit schematic of a single receive element on the ex vivo sample receive coil, containing the balun circuit, split capacitors, active detuning and preamplifier attached using a λ/4 coaxial cable. (E) Isometric (left) and side views (right) of a SolidWorks render illustrating the placement of the 16 receive channels with respect to the sample.

For the sample housing, a cylindrical container was designed in SolidWorks (Dassault Systèmes SolidWorks Corporation, MA, USA) and 3D printed using SOMOS XC1112 (DSM, Heerlen, NL) material (Fig. 1C). This particular material was chosen as it possesses susceptibility close to that of water, helping us reduce susceptibility-induced artifacts. The container was designed to have an outer diameter of 82 mm and a wall thickness of 1 mm, allowing for an internal diameter of 80 mm. With a length of 90 mm, the sample container was able to accommodate large post-mortem brain samples such as the single human occipital lobe that was used here (see below). The container was closed using a 5 mm thick 3D printed cover with a sealing groove along its circumference. As sealant either silicon pressure paste (Dow Corning, Seneffe, BE) or standard water-proof quick drying silicone sealant was applied in the groove of the container cover which was then pressed down onto the container to provide a water-tight fit. A hole was drilled onto the container cap in order to fashion a degassing port, which was then plugged using a nylon screw and a nitrile O-ring. Additional Kapton tape was applied to the outside of the container to secure the cover and provide a tight fit in the coil former. A 3D CAD layout showing the distribution of the 16 receive coils across the sample, as rendered in SolidWorks, is shown in Fig. 1E.

RF-coil circuitry

Each receive coil was connected to a lattice or LC balun circuit (Reykowski et al., 1995, Wiggins et al., 2006) and a tuning and matching network comprising of high-voltage ceramic trimmer capacitors CT and CM (1–10 pf, Johanson Technology, CA, USA) for tuning the coil to resonance at 9.4 T and matching the coil's output impedance to 50 Ω. Passive detuning circuitry consisting of an LC circuit in parallel with a high voltage diode was placed, along with the balun circuit, on small, flexible substrates (Rogers RT5870; Rogers Corporation, Gent, BE). All coils were tuned and matched to 400 MHz and connected to low input-impedance preamplifiers (WMA9RA, WanTcom Inc., MN, USA) using λ/4 (130 mm) length coaxial cables (Huber-Suhner, K02252D), in order to achieve preamplifer decoupling between individual coil elements - by transforming the high impedance at the coil output to the corresponding low impedance at the preamplifier input (Roemer et al., 1990, Reykowski et al., 1995). An active detuning circuit was fashioned on the balun circuit, using the balun inductor and a PIN diode across the matching capacitor. When forward biased, the parallel resonant LC circuit adds a high impedance in series with the coil loop, blocking current flow in the receive loop during transmit.

The preamplifiers were arranged in a circular layout mirroring the receive coil layout and are placed on a circular preamplifier motherboard which contains the required circuitry for providing the PIN diode voltages for actively detuning each receive coil element during transmit mode. In this layout, the preamplifiers were arranged along the B field or the z-direction of the magnet, thus minimizing any Hall effects which might affect the field effect transistors (FETs) used in these preamplifiers (Keil and Wald 2013). Cable traps on the output of each preamplifier were required and implemented – by fashioning a solenoid (2 turns, 4 mm diameter) from the coaxial preamplifier output cable and bridging a variable capacitor across its ends and tuning it to the resonant frequency of 400 MHz, all positioned outside the immediate RF receive loop locations. The traps helped to significantly reduce common-mode currents and suppressed interactions with the transmit coil.

For transmit, a separate 16 channel pTX transmit coil (Shajan et al. 2014) was used. The receive coil assembly was designed to be concentric to the transmit coil in order to allow for easy integration between transmit and receive sections (Fig. 1B).

Sample preparation

Tissue used for imaging was obtained post mortem from the left hemisphere of a male subject, without known neurological or psychiatric disorders. The tissue donor gave his informed and written consent to the donation of his body for teaching and research purposes regulated by the Dutch law for the use of cadavers for scientific research and education. Accordingly, a handwritten and signed codicil from the donor, posed when still alive and well, is kept at the Department of Anatomy and Embryology Faculty of Health, Medicine and Life Sciences, Maastricht University, Maastricht, The Netherlands. The approximately 80×80×80 mm3 formalin-fixed occipital lobe of the left hemisphere was first immersed in 6 times its volume of phosphate-buffered solution (PBS) for 6 weeks to wash out formalin and rehydrate the sample. After washing, the sample was placed in the 3D printed cylindrical container and embedded in proton-free, susceptibility matched fluid Fluorinert FC-3283 (3 M, Delft, NL) for imaging. The 3D printed container provides a sample holder susceptibility-matched to the imaging medium and the sample itself. The tight fitting cylindrical container fits the receiver coil and can accommodate other samples, which fit the 80 mm diameter cylindrical volume. Designing a cylinder and filling it with proton-free susceptibility-matching fluid was found to be simpler and more flexible than constructing a conformal sample-holder for each sample. The container was then sealed and the cover pressed down to allow the silicone to distribute evenly along the lid. Using a syringe, Flourinert was injected into the cylinder to replace air bubbles and the degassing port sealed.

RF-coil bench measurements

Bench tests were performed using a Agilent HP E5071C ENA Series network analyser, a customized coil-plug bed and a cylindrical phantom (49.8% demineralized water, 48.8% sucrose, 1.3% KCl, 0.10% Dowicil; Max Planck Institute for Biological Cybernetics, Tübingen, Germany) as load. Measurements for loaded (QL) and unloaded (QU) Q ratios, receive coil decoupling and preamplifier decoupling for each coil element were performed. The bench tests were undertaken to validate coil tuning and matching, preamplifier decoupling for each coil element and active coil detuning.

The coil quality factor ratio (QU/QL) for a single receive coil element was measured using a dual-loop decoupled (~50 dB) inductive probe coupled to a single Rx element: first as a single isolated element outside the receiver array and then within the populated array while keeping all other receive elements in a detuned state. Each loop on the receive array was tuned to the Larmor frequency of 400 MHz and matched to 50 Ω. Coupling between neighbouring elements was measured through S21 measurements by connecting the coil outputs to the network analyser, keeping all other elements detuned. Using this method, the overlap between neighbouring elements could be further optimized to ensure minimal mutual inductance.

Preamplifier decoupling of a single loop was measured as the difference in S21 measurement using a pair of decoupled pickup probes, first with the coil power matched to 50 Ω under load but without the preamplifier present and then with the coil terminated using the low input impedance preamplifier, while keeping all other coils in the receive array detuned. The active detuning for each receive element was measured as the difference in an S21 measurement, between when the coil is matched to 50 Ω and when detuned.

MRI acquisition and analysis

Data acquisitions were performed on a 830 mm human size bore research 9.4 T Siemens MAGNETOM MR system (Siemens Healthcare, Erlangen, Germany) interfaced with a 16Ch parallel transmit (pTx) coil. Prior to acquiring high resolution anatomical data, B and B+ shimming were performed. First, a localizer was acquired for spatial-localization reference; then a fieldmap (Cusack and Papadakis 2002) and DREAM (Nehrke and Bornert 2012) protocol were acquired to characterise B and B+ map profiles, respectively, along the sample. MATLAB (MathWorks, MA, USA) routines were used to optimise B and B+ shimming using those acquisitions (Setsompop et al., 2008, Tse et al., 2016). Fieldmap sequences were acquired and shimmed iteratively, two to three times, to improve B homogeneity.

All acquisitions were performed with a custom 3D gradient echo (GRE) pulse sequence, modified to allow large matrix sizes and to use a composite parallel excitation pulse using the kT-points technique for B+ homogenization (Cloos et al. 2012). An Actual Flip-angle Imaging (AFI) sequence (Yarnykh 2007) was acquired later to verify B+ homogeneity. The 3D GRE sequence was used to acquire high resolution isotropic T2* weighted data at: 500 μm (Repetition time (TR)/Echo time (TE)=36 ms/17.68 ms, flip angle (FA)=35deg, readout bandwidth (BW)=220 Hz/px, matrix dimensions=160×156×160, acquisition time=00:11:45), 100 μm (TR/TE=47 ms/22ms, FA=49deg, BW=30 Hz/px, matrix dimensions=800×800×832, acquisition time=06:49:29) and 60 μm (TR/TE=46 ms/20.1 ms, FA=49deg, BW=40 Hz/px, matrix dimensions=1400×1400×1280, acquisition time=17:59:00) isotropic resolution. A multi-echo 3D GRE protocol was used using monopolar RO-gradient to acquire multi-contrast T2* weighted data for the purpose of quantitative T2* mapping at 200 μm isotropic resolution (TR/TEs=38.5 ms/6.45, 14.69, 22.93 and 31.17 ms, FA=25deg, BW=220 Hz/px, matrix dimensions=400×400×416, acquisition time=01:24:56. The in-plane field of view (FoV) for all the experiments was 80×80 mm. For all the acquisitions, a 0 V transmit power noise reference scan was acquired with matching protocol parameters (e.g. bandwidth, FoV and base resolution). Raw data was streamed off the MR system in realtime for later image reconstruction. Offline reconstruction was performed using home-made routines in python and MATLAB. No filtering was applied in the reconstruction to show the raw quality and effective resolution and signal-to-noise ratio (SNR) of the acquired data. For high resolution GRE, T2* weighted anatomical volumes Roemer reconstruction (Roemer et al. 1990) was used and for quantitative T2* mapping a noise covariance corrected root of sum of squares (cov-rSoS) reconstruction (Triantafyllou et al. 2011) was used.

To compare the SNR of the ex vivo sample coil with a more widely available alternative, equivalent data was acquired in the same scanner system using a 16 channel transmit, 31 channel receive whole-head coil (Shajan et al. 2014). The excitation profiles from both measurements were acquired using the AFI sequence in pTx mode using kT-points (in the dedicated 16 channel receive coil) and AC shim (in the Head coil). A proton-density weighted GRE using monopolar readout at 1000μm isotropic resolution (TR/TE=1000 ms/4.0 ms, FA=88deg, BW=300 Hz/px, matrix dimensions=82×90×80) was used with a matched zero volt noise acquisition as described above. Image SNR for both coils was computed using a pseudo-replica approach (Robson et al., 2008, Shajan et al., 2016) with the subsequent noise data. We generated 100 replicas by adding random samples from the noise scan onto the GRE k-space of the standard image acquisition prior to Fast Fourier Transform (FFT), while maintaining the noise covariance between the receive channels. The final image SNR was then calculated as the ratio of the mean of the image to the standard deviation of the noise over these 100 replicas. Resulting SNR maps for the ex-vivo sample coil and for the Head coil were corrected by a normalized B+ map calculated from AFI acquisition using kT-points. Coregistration between AFI and GRE volumes were performed before B+ calculation using SPM 12. This corrects any underestimation of SNR values caused by B inhomogeneity leading to flip angles away from Ernst angle.

Quantitative T2* mapping was performed in a voxel-based analysis approach using the GPU accelerated python-based Maastricht Diffusion Toolkit (Harms et al. 2015) by Markov Chain Monte Carlo (MCMC) sampling. To this end a mono-exponential decay equation including an S0 term was defined in MDT's OpenCL-based signal modelling language. MCMC sampling was performed using the random walk Metropolis algorithm (Peters et al. 2007) with a uniform prior for S0 and T2* with lower and upper bounds of [0 to 50] a.u. and [0.0, 100.0] ms respectively and an offset-Gaussian likelihood function to account for the Rician rectified noise floor. The convergence of the chains and unimodality of the sampled posterior was visually inspected for several voxels leading to the following settings: 1500 burn-in samples, 100 samples with a step of 5. Gaussian distributions were fitted to the histograms of the final samples to obtain the mean (max) posterior estimate and the standard deviation of the posterior (MacKay et al., 2006, Behrens et al., 2007, Orton et al., 2014). The log of the standard deviation is reported as a measure of the uncertainty (variability, inverse precision) of the S0 and T2* maximum posterior estimates. A 3D surface reconstruction of the entire left occipital lobe was constructed in BrainVoyager QX 2.8.4 (Maastricht, The Netherlands) based on the 200μm qT2* maps. Region growing operations, followed by erosion, smoothing and then dilation steps were used to create a surface at the approximate pial surface without contamination of vessels and pia mater to visualize the macro-anatomy of the sample (Goebel et al., 1998, Kriegeskorte and Goebel, 2001, Goebel et al., 2006, Frost and Goebel, 2012).

Results

RF coil characterization

The unloaded Q factor QU (indicative of coil losses) for an isolated receive element of the 16 channel ex vivo sample coil at 400 MHz, was about 248 and the loaded Q factor QL, indicative of coil and tissue losses, was 38. For a single receive loop surrounded by 5 detuned loops in the array, the same Q factor measurement yielded a QU/QL of 245/42. All receive coils were matched between −21 dB and −28 dB for the occipital lobe sample. The active PIN diode detuning provided an isolation better than 28 dB between tuned and detuned states. The decoupling between neighboring, overlapping receive elements ranged between −14 dB to −19 dB. Decoupling between next-nearest neighboring elements (or non-overlapping elements) of the array ranged between −17 dB to −25 dB. Preamplifier decoupling achieved an additional 20 dB of isolation. The noise correlation matrix, as shown in Fig. 2A, was obtained with ex vivo scanning, with the correlation ranging between 0.2% and 51% with an average of 22%.

Fig. 2

(A) Noise correlation matrix for the 16 channel receive coil with the scale normalised to 1 (B) Individual coil sensitivity profiles for the coil, with the slices shown aligned along the center of the coil elements. 100 μm isotropic GRE image using a (C) cov-rSoS reconstruction and using a (D) Roemer reconstruction respectively.

Fig. 2 shows the receive characteristics of the ex vivo sample coil. Fig. 2A shows the noise correlation matrix with a maximum correlation of 0.51 and an average of 0.22 and Fig. 2B shows the single channel images for the occipital lobe sample across the two rows of coils (clockwise across each row) with high SNR near the coil surface showing good delineation of WM/GM. Figs. 2C and 2D show coil-combined reconstructions for a 100 μm isotropic GRE acquisition along a central coronal slice using a cov-rSoS reconstruction and Roemer reconstruction respectively. Substantial receive penetration to image the entire sample effectively can be seen, especially when using a coil-sensitivity corrected Roemer reconstruction (Fig. 2D).

Fig. 3 shows a comparison between the transmit profiles at 9.4 T single channel CP mode and 9.4 T parallel transmit (pTx) using the kT-points method. The 9.4 T CP excitation profile is less homogeneous with greater central brightness in the coronal view and a sharp falloff of flip angles from the posterior to the anterior end. However, the use of kT-points at 9.4 T homogenizes the excitation profile along the sample, decreasing the over-excitation in the posterior part (from 50% to almost 10%) and increasing in the anterior part (from −25% to −5% in average).

Fig. 3

Normalized transmit (B+) profile maps of the 9.4 T CP (first column) and 9.4 T kT-points homogenized (second column) acquisitions. Upper row shows a near sagittal slice with the dotted white line in the first row indicating the coronal slice shown in the second row.

Fig. 4 shows a comparison of image SNR between the two coils for a 1mm isotropic proton-density weighted GRE acquisition. The measured uncorrected SNR (top row) shows a drop-off in SNR away from the receive coils (i.e. peripheral to central) across both acquisitions. Nonetheless, the SNR with the dedicated 16 channel coil is considerably higher across the entire volume than in the respective acquisition using the head coil, at the periphery by about a factor of five and at the center by a factor of almost 2. The B+ corrected SNR (bottom row) shows an estimate of the theoretical SNR that would be achieved under completely homogenous excitation at the Ernst angle (close to 90deg for long TR PD weighted). This shows higher SNR at the center and periphery compared to the uncorrected SNR, where the kT-point homogenized pulse still slightly overflips (in the center) or underflips (in the periphery) with respect to the Ernst angle.

Fig. 4

SNR maps for the 16 channel dedicated ex-vivo sample receive coil (left column) and the whole head coil (right column) for a proton-density weighted GRE acquisition, without B+ correction (top row) and with B+ correction (bottom row).

Anatomical and quantitative T2* imaging

Fig. 5 shows successively zoomed transverse views of two T2* weighted GRE acquisitions of the occipital lobe sample at 100 μm and 60 μm isotropic resolutions. No spatial filtering was used in the image reconstruction to present the quality of the raw data. The cut plane through the sample is slightly different between acquisitions as they were acquired during different sessions and oblique views, which would interpolate the data, were avoided for an accurate comparison. The zoom-ins on the right show the same approximate location in the sample. The anatomical detail visible (along the calcarine sulcus) across both resolutions shows the stria of Gennari (SoG) in the middle of the cortical ribbon, with a tendency of the 60 μm images to show better delineation of the stria. There is also a tendency for a better definition of the pial boundary in the higher resolution 60 μm dataset. However, the middle upper panel for both A & B shows that in T2* weighted images, contrast can vary with location and becomes dependent on global contrast scaling. For instance, the nearly horizontal bank above the red box may have better defined stria for the 100 μm dataset in the current view, because contrast was adjusted for each of the 100 μm and 60 μm individually to the white box. This shows the need for quantitative contrast which does not vary with position. The red zoom-in inset details a vessel wall which shows how an even better definition of high-contrast structures can be achieved at 60 μm resolution. However, more than identifying an optimal resolution, these results show that cortical laminar structures are clearly depicted at high resolution of 100 μm and above, and high SNR.

Fig. 5

Transverse slices of GRE acquisitions for (A) 100 μm isotropic (B) and 60 μm isotropic of the occipital lobe sample at different zoom levels individually contrast adjusted.

Fig. 6 shows that quantitative T2* (qT2*) mapping can be performed at 200 μm isotropic resolution for the entire occipital lobe using MCMC sampling. Fig. 6A shows sagittal views through the cov-rSoS reconstructed single echo datasets where the receive profile inhomogenities are evident. Figs. 6B and 6C show the estimated S0 and qT2* maps, respectively, averaged over the different FA acquisitions. The S0 map absorbs proton density and non-quantitative transmit and receive inhomogenities. A good T2* estimate is achieved over the whole sample with clear delineation of white matter and gray matter and a higher T2* in the superficial gray matter layers than in the deeper layers. The log standard deviation of the posterior distribution of the MCMC chain (right images in Figs. 6B and 6C) is an indication of the uncertainty of the estimates. The uncertainty of the T2* estimate is slightly higher in gray matter than in white matter and clearly lower close to the coils where signal is higher (lower left) and higher far away from the coils (on the right). The uncertainty of the S0 estimate is higher in white matter. Fig. 6D shows a histogram of the T2* values observed in Fig. 6C, with small peaks corresponding to the white matter T2* of approximately 20ms and a deep gray matter T2* of approximately 26ms.

Fig. 6

Quantitative T2* mapping of the entire occipital lobe at 200 μm calculated from a multi-echo GRE acquisition using 4 TEs. (A) Individual echoes, (B) S0 estimate, mean (left) and log standard deviation (right) and (C) T2* estimate, mean (left) and log standard deviation (right) maps. (D) A histogram of T2* along the same slice is shown where small peaks (red arrows) correspond to white and gray matter T2* values.

Fig. 7A shows a 3D surface reconstruction of the entire occipital lobe using the 200 μm qT2* map, showing the extent and macro-anatomical features of the sample. The four coloured lines indicate the anterior-posterior (AP) locations along the calcarine sulcus (CS, indicated by arrows) for which coronal slices through the qT2* map are shown in Fig. 7B and zoomed views of the CS are shown in Fig. 7C. The SoG is clearly visible along the cortex in the depth of the CS at all AP positions and, with some scrutiny, can be followed at most locations to its termination on gyri encroaching the CS at the putative border between primary (functional area V1, Brodmann area 17) and secondary (functional area V2, Brodmann area 18) visual areas. The distinctive patterns of a higher T2* in the superficial gray matter layers than in the deeper layers, separated by the SoG, is visible everywhere in putative V1. Finally, it is interesting to note the qT2* contrast visible in parts of white matter, particularly the lower T2* in parts close to the lateral ventricle, tentatively corresponding to large myelinated tracts.

Fig. 7

3D surface reconstruction of entire left occipital lobe based on 200 μm qT2* maps (A) showing a medial view (left) and lateral view (right). (B) Coronal slices at four different anterior posterior positions along the calcarine sulcus, shown in different colours (C) and zoomed-in views of anatomical details along the same slices showing layer definitions and fiber tracts in white matter. CS: Calcarine sulcus (indicated by coloured arrows). SCC: Splenium of the Corpus Callosum.

Discussion

We constructed a 16 channel cylindrical phased array receive coil to image a large post mortem human occipital lobe sample (~80×80×80 mm3) in a wide-bore 9.4 T human scanner. The achieved SNR is strongly increased compared with a human in-vivo head coil at 9.4 T. Although the transmit profiles with a circularly polarized (CP) transmit mode at 9.4 T is quite inhomogeneous over the large sample, this challenge is successfully resolved with parallel transmit (pTx) using the kT-points method. Increased spatial definition of cortical details is visible in 60 μm anatomical images for the entire occipital lobe compared to lower resolutions. In addition, we were able to achieve sufficient control over SNR, B and B homogeneity and multi-contrast sampling to perform quantitative T2* imaging over the same volume at 200μm.

9.4. T large post mortem sample RF-coil

The constructed dense, cylindrical 9.4 T phased-array receive coil was targeted at acquiring high-resolution anatomical and quantitative MRI of large post mortem samples. The coil's performance was evaluated through bench tests and ex vivo human occipital lobe imaging. The placement of the 2 rows of receive arrays covers the entire occipital lobe sample, but also ensures that the small diameter coils are as close as possible to the region of interest. This greatly enhanced their receive sensitivity and achieved SNR when compared to that of a larger loop 31Ch receive, 16Ch transmit head coil at 9.4 T. The image SNR measured shows a five-fold increase in maximum SNR (peripherally at the cortex) for the ex vivo sample coil over the compared head coil. This can be attributed to the smaller coil loops in close proximity to the sample, enabling much higher SNR in those areas. With the head coil, only a few receive elements would couple with the sample (due to its size and positioning in the coil), thus reducing overall SNR in areas further away from coil elements. The position and placement of the coil elements in the 9.4 T ex vivo sample coil allows for higher SNR throughout the sample although the gain is proportionally higher at the periphery, close to the coils, as is evident from Fig. 4, although SNR is still better by a factor of two in the center. For the occipital lobe sample this provided very high SNR specifically in most of the cortex and superficial white matter.

The presented ex vivo sample coil layout could be implemented at 7 T, in which case it would also be expected to give significant gains for samples that fit its size over e.g. a commercial head coil. This investment can be worthwhile in a situation of regular large post-mortem sample experiments. The up to five-fold SNR per unit time efficiency increase makes available a level of resolution and quantitative contrast for large samples that is otherwise not obtainable, unless at an order of magnitude longer acquisition times which in the long run is potentially more expensive and time consuming. While the 7 T field strength would yield more homogeneous B and B transmit fields than at 9.4 T (and less requirement for parallel transmit), it would also result in 33% less overall SNR and lower T2* and other magnetic susceptibility moderated contrast, such as phase contrast and quantitative susceptibility imaging (QSI) (Augustinack et al., 2005, Fukunaga et al., 2010, Lee et al., 2012). The post mortem sample size of 80×80×80 mm3 targeted here would not be trivially accommodated in preclinical or animal MRI systems. Current preclinical systems with wider bore diameter magnets will generally be able to accommodate the sample using their largest diameter (and lowest performing) gradient and animal body RF coils. Although such gradient coils would still outperform large human bore gradients, it should also be noted that the GRE imaging reported here is not particularly gradient-performance limited. Furthermore, imaging with smaller preclinical gradients would likely be affected by reduced homogeneity in main B field and linearity in gradient fields over the full 80×80×80 mm3 FoV, due to the smaller diameter of gradient and shim coils. Most importantly, an insufficient receive channel count and clear bore space would likely be unable to support the customised 16 channel RF coil (along with a separate transmit coil) which provides most of the SNR gains reported. Having said this, we should note that the main conclusion to draw from the reported results is not that large human bore systems are superior to smaller bore preclinical systems for high resolution imaging of brain samples of this size. Rather, the results support the notion that phased-array receive coils specifically constructed for post mortem sample imaging can provide significant gains over more generic solutions, which can be generalized to even larger samples, such as entire human post mortem brains, on large bore systems.

The B+ transmit profiles in CP mode for 9.4 T show the expected high central fields dropping off towards the periphery since the sample dimensions are close to the wavelength at 9.4 T. In addition, there is asymmetry along the anterior-posterior direction with the transmit field dropping off to the anterior sample side in the acquired orientation. However, this issue is largely rectified by using kT-points which results in a more homogeneous excitation profile along our region of interest. For the T2* weighted GRE imaging, and especially the qT2* analysis, performed here, the influence of this B homogenization on contrast is relatively low. However, this additional control over B can also be of some influence on SNR in T2* GRE imaging as the B inhomogeneity corrected SNR maps show. More concretely, comparing corrected to uncorrected SNR show how much of theoretically achievable SNR is actually realized with kT-point homogenization. In short, B transmit field homogenization is a tool in this case to help realize the full SNR gain from the receive side RF-coil. To ensure an even more homogeneous excitation profile, future work could involve building a smaller diameter, sample-specific transmit coil with pTx capabilities at 9.4 T. However, in this work we chose the sample size as similar to (and a bit under) the RF wavelength at 9.4 T, which is about 9 cm, which leads to relatively benign B+ inhomogeneity effects. This creates a good starting position for pTx enabled B+ homogeneity, which can be achieved by both large and small Tx coil configurations. Therefore the gain of constructing a smaller diameter transmit coil would be less than for the step from CP to kT-points on the larger diameter Tx coil.

High resolution anatomical and quantitative MRI

The high Rx signal allowed us to acquire 100μm and 60μm isotropic T2* weighted data over the entire sample. While the SoG is clearly evident in both the 100 μm and 60 μm datasets, it is slightly better spatially defined in some regions the 60 μm scan. The higher contrast vessel wall and outer cortical boundary even more clearly show the better spatial definition which can be achieved. However, contrast in the T2* weighted images can vary with location and becomes dependent on global contrast scaling, which shows the need for quantitative contrast which does not vary with position and is homogenous across the entire sample. Although no spatial filtering was applied in the image reconstruction to present the quality of the raw data, a small degree of k-space filtering may be preferred in future neuroanatomy studies to mitigate a slightly visible Gibbs ringing.

The qT2* mapping of the entire human occipital lobe at 200 μm isotropic shows some of the advantages of quantitative imaging. These include an even clearer definition of cortical architecture details, specifically here the SoG, inherent control of low frequency contrast inhomogeneity without a need to estimate receive coil sensitivity, and generally comparable images in quantitative physical units. The estimated S0 shows a signal profile similar to the data absorbing information about the proton density as well as receive coil sensitivity. In our sample, T2* values in GM are greater than in WM and are also greater in the superficial GM layers than in the deeper layers. T2* relaxation time is inversely proportional to the concentration of myelin (and iron content), which is highest in WM and higher in the deeper layers of the GM. The calculated T2* values are higher than those acquired with in vivo subjects (Peters et al. 2007). This could be partly explained by the hydration state of the tissue and the lack of active venous vasculature which would otherwise contribute to lower T2*. The estimated T2* values are also comparable with previously reported T2* values in post-mortem samples (Reeves et al. 2016). T2* estimates can be affected by residual static B inhomogeneity due to imperfect B shimming (Fernandez-Seara and Wehrli 2000). This seems to be visible in Fig. 6C where we notice slightly less contrast in a small area in the posterior-most section. Further improved iterative shimming routines could resolve this in the future. In addition, the rectified Rician noise floor can have an effect on the fitting of long TE low signal echoes. However, SNR was sufficiently high even in anterior parts of the sample, such that T2* could still be estimated, albeit with greater uncertainty. Moreover, MCMC sampling was performed with an offset-Gaussian likelihood function to account for the rectified noise floor.

The 200 μm qT2* imaging allows mapping of the SoG along the entire length of the calcarine sulcus. The SoG is clearly visible along the cortex in the depth of the CS at all AP positions in Fig. 7 and can be followed at most locations to its termination at the border between primary (functional area V1, Brodmann area 17) and secondary (functional area V2, Brodmann area 18) visual cortex. This illustrates the capacity to map out and atlas architectural area boundaries in future studies.

Outlook

Future work can involve quantitative imaging of T1, T2, phase and magnetic susceptibility in addition to T2*. T1 imaging could be realized with relaxometry on mixing times in STEAM imaging or fitting of multiple flip angles in GRE imaging (Lutti et al. 2014). However, both would place an even higher demand on B+ homogeneity and control and quantification of absolute flip angles. Phase and quantitative susceptibility mapping (QSM) would require reconstruction, unwrapping and fitting of phase instead of magnitude data (Deistung et al. 2013). Quantitative MRI (qMRI) at resolutions higher than 200 μm, up to 100 μm isotropic, may be possible using the current setup. In our experience, challenges here lie more in processing, analysing and visualizing the data than in the acquisition, as multi contrast 100 μm acquisition could in principle be performed. qMRI at even higher resolutions up to 60 μm isotropic might not be feasible due to insufficient SNR and multi contrast sampling leading to very long acquisition times. Diffusion MRI (dMRI) is another interesting quantitative technique to extend to large samples at very high resolutions. Post mortem dMRI has been performed previously over small FoV (Dell'Acqua et al., 2013, Kleinnijenhuis et al., 2013, Leuze et al., 2014, Aggarwal et al., 2015, Modo et al., 2016) and allows in situ validation by histology measures (Leergaard et al., 2010, Axer et al., 2011, Budde and Frank, 2012, Seehaus et al., 2013, Magnain et al., 2014, Khan et al., 2015, Seehaus et al., 2015). Diffusion MRI is inherently more SNR challenged than GRE imaging and benefits less from moving to UHF. In addition, it also requires high flip angle pulse sequences and good B+ homogeneity, making it a challenging prospect. Finally, the platform for UHF cortical layer specific quantitative MRI can be extended to even larger FoV allowing larger samples up to the entire post mortem human brain.