Non-BOLD contrast for laminar fMRI in humans: CBF, CBV, and CMR₀₂

Abstract

Functional magnetic resonance imaging (fMRI) using the blood oxygenation level-dependent (BOLD) contrast indirectly probes neuronal activity changes via evoked cerebral blood flow (CBF), cerebral blood volume (CBV) and cerebral metabolic rate of oxygen (${\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$) changes. The gradient-echo BOLD signal is mostly sensitive to ascending veins in the tissue and to pial veins. Thereby, the achievable spatial specificity to neuronal activation is limited. Furthermore, the non-linear interaction of CBF, CBV and ${\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$ can hamper quantitative interpretations of the BOLD signal across cortical depths with different baseline physiology. Measuring CBF, CBV or ${\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$ directly on a depth-dependent level has the potential to overcome these limitations. Here, we review these candidates of physiologically well-defined contrasts with the particular focus on arterial spin labeling (ASL), vascular space occupancy (VASO) and calibrated fMRI. These methods are reviewed with respect to their fMRI sequence parameter space and the applicability for neuroscientific studies in humans. We show representative results of depth-dependent ‘non-BOLD-fMRI’ in humans and their spatiotemporal characteristics. We conclude that non-BOLD methods are promising alternatives compared to conventional fMRI as they can provide improved spatial specificity, quantifiability and, hence, physiological interpretability as a function of cortical depth. At submillimeter resolution with inherently low signal-to-noise ratio (SNR), however, their use is still challenging. Nevertheless, we believe that ‘non-BOLD-fMRI’ is a useful alternative for depth-dependent investigations, by providing valuable insights into neurovascular coupling models that facilitate the interpretability of fMRI for neuroscientific applications.

1. Introduction

The cortical grey-matter (GM) ribbon consists of six histologically-defined layers with thicknesses between 0.2 and 1 mm. Mapping neural activity across thin cortical layers and their differential functional connectivity to distant brain areas is highly valuable for cognitive neuroscience, and the required methodology has become the focus of cutting-edge research. The only non-invasive in vivo method currently capable of mapping brain activity at submillimeter resolution and, hence, having the potential to detect layer-dependent physiological changes, is functional magnetic resonance imaging (fMRI), such as the widely applied blood oxygenation level-dependent (BOLD) fMRI. The BOLD-relevant oxygenation changes in draining veins (Menon et al., 1995; Turner, 2002) result, however, in a reduced effective spatial resolution, compared to the nominal imaging resolution. Aside of local specificity limitations in BOLD-fMRI, its quantitative interpretation of depth-dependent signal change can be hampered by heterogeneous baseline physiology, such as baseline venous blood volume. To account for these challenges in depth-dependent BOLD-fMRI, alternative methods have been proposed. Popular ‘non-BOLD-fMRI’ contrasts at standard spatial resolutions are sensitive to changes in cerebral blood flow (CBF), cerebral blood volume (CBV) and cerebral metabolic rate of oxygen (${\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$).

In this article, we first provide an overview of the methods to measure these physiological parameters. The currently most promising methodologies in human application are identified to be arterial spin labeling (ASL) for CBF, vascular space occupancy (VASO) for CBV and calibrated fMRI (a.k.a. ‘calibrated BOLD’) for ${\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$. Their experimental setups and contrast mechanisms are summarized and compared to conventional gradient-echo (GE)-BOLD-fMRI. Representative depth-dependent results are shown. Subsequently, we discuss the advantages and challenges of the respective methods considering their sequence requirements and confounding factors. In light of this discussion, we sought to provide recommendations of appropriate sequence parameters for cortical depth-dependent applications. Finally, all methods are directly compared with regards to their sensitivity, specificity, sampling efficiency, quantifiability, availability and invasiveness.

Cortical depth-dependent non-BOLD-fMRI approaches were initially established in animal models. In this review, however, we are mostly focusing on their current and potential applications in humans. In the introductory part of the article, we provide a brief overview of the available non-BOLD imaging methods, including reviewing and referencing pioneering animal studies. However, the presented data, the discussion of the advantages and disadvantages of non-BOLD fMRI, and the recommendations of best practices of non-BOLD depth-dependent fMRI are focusing on non-invasive methods for applications only in humans.

Aside of $\mathrm{C}\mathrm{B}\mathrm{F}-,\mathrm{C}\mathrm{B}\mathrm{V}-,{\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$-based fMRI methods discussed in this article, there are other potential substitute methods for depth-dependent fMRI, such as spin-echo (SE) BOLD fMRI (De Martino et al., 2013) or diffusion-weighted fMRI (Truong and Song, 2009), which are not discussed in this article.

2. Cerebral blood flow

2.1. Definition of CBF

Cerebral blood flow is a key physiological parameter indicating microvascular density during rest, and indicating vascular reactivity and strength of stimulus processing during functional activation. Indirectly, resting CBF reflects brain metabolism indicated by a relative homogeneous oxygen extraction fraction (OEF) despite large variation in CBF over the whole cortex (e.g., (Gusnard and Raichle, 2001)). How much functional CBF changes are proportional to changes in oxidative metabolism is a matter of debate. In fact, different coupling constants $n=\mathrm{\Delta }\mathrm{C}\mathrm{B}\mathrm{F}/\mathrm{\Delta }{\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$ have been determined, using the calibrated fMRI approach (see section 4 below), for different brain areas and even for different tasks in the same brain area ((Buxton et al., 2014) and references therein).

CBF is typically defined as the amount of blood delivered to a tissue per unit of time and tissue mass, that is, the unit is ml/100g/min. As the average brain tissue density is approximately 1 g/ml, for a typical CBF value in the human cortex of 60 ml/100g/min, we obtain 0.01/s indicating that after 1s, 1% of the tissue is composed of inflowing blood from outside the respective tissue volume.

2.2. Methods for measuring CBF

There are many invasive and non-invasive methods to measure CBF, and they can be roughly classified into two categories:

1. With ‘intravascular’ contrast mechanisms, the passage of either an exogenous contrast agent (CA) or red blood cells (RBC) within the blood vessels is monitored and, under some biophysical assumptions, converted to CBF. The two measures only yield identical results assuming proportionality between RBCs and plasma density. In these intravascular approaches, special care has to be taken to only probe capillary signal changes, as the acquisition methods themselves do not distinguish between different vascular compartments.

   The choice of CA depends on the imaging method. For MRI, CAs altering the longitudinal ($T_1$) and/or effective transverse relaxation time ($T_2^{\mathrm{*}}$), such as GdDTPA²⁻, are injected into a vein. Some of the CAs are approved for human use and can be utilized in clinical populations.

2. With ‘diffusible tracer’ contrast mechanisms, substances that pass the blood-brain barrier and exchange (in the capillaries) with the brain tissue are measured. Thus, the acquisition method, if properly executed, yields results that adhere to the above definition of CBF as reflecting capillary plasma flow. However, artifacts due to larger vessels (e.g. arteries) can be present for short transit time delays, which must be carefully avoided in order not to bias the CBF values and their spatial distribution. The gold standard diffusible tracer contrast is to use a radioactively labeled CA and autoradiography to count the number of radioactive events in the local tissue representing CBF. The imaging parameters must obey certain constraints, such as complete delivery of the CA bolus to the tissue (see below for ASL), and model assumptions have to be made to convert signal changes into CBF values ((Buxton, 2005) and references therein).

Functional CA method require a BOLD EPI type acquisition scheme with the same spatial and temporal resolutions that can be easily met with other modern non-CA fMRI methodology. In order to estimate voxel-wise CBF values, a time course of signals and an accurate estimate of the arterial input function to the local region is needed to conduct a deconvolution analysis. This can be cumbersome when CA methods are used for functional experiments.

For laminar fMRI, (intra-vascular and diffusible tracer) CA methods cannot be routinely utilized as they are limited by long acquisition times. An additional limitation for applications of CA methods in fundamental neurosciences in healthy subjects might be ethical concerns. As soon as layer-fMRI applications are developed from which clinical populations could benefit, we believe that layer-dependent functional CA applications might become more popular.

2.2.1. Arterial spin labeling MRI

The most promising approach in humans for assessing laminar CBF is ASL (Alsop et al., 2015; Calamante et al., 1999). It is non-invasive and utilizes radiofrequency (RF) pulses to label inflowing water in the arterial blood (Fig. 1A). The resulting signal intensity in each voxel is then composed of the signal intensity of the static tissue reduced by the longitudinal magnetization of the inflowing blood. In order to determine and remove the contribution of the static tissue, a separate image is acquired without the labeling RF pulse. Variants of ASL methodology are distinguished by how the labeling is performed and magnetization and vascular artifacts are controlled.

Figure 1: CBF estimation with ASL.

The contrast-generation mechanisms of ASL is schematically illustrated in panel (A). Results in panels (B-E) were acquired with 3D-EPI FAIR QUIPSSII at 7T by Dimo Ivanov as described in (Ivanov et al., 2016b). Panel (B) depicts representative maps of perfusion signal at 0.7mm isotropic resolution, and the corresponding tSNR. In the map of CBF-weighted signal, it can be seen that the voxels with highest perfusion-weighted signal align along the cortical ribbon (blue insert). Panel (C) depicts the corresponding cortical profiles in the visual cortex. Note that the perfusion signal is largest at middle cortical depths. The relatively low CBF signal at the cortical surface indicates that CBF-fMRI is not biased by large pial veins as compared to conventional GE-BOLD. (D-E) depict the functional results of a flickering checkerboard paradigm acquired at 0.8mm isotropic resolution. It can be seen that BOLD is largest at superficial depths, while most of the CBF activation is located at deeper cortical depths (green arrows).

The main ASL approaches are: pulsed ASL (PASL), continuous ASL (CASL) and pseudo-continuous ASL (pCASL) ((Alsop et al., 2015) and references therein). In PASL, an RF pulse labels water protons in a slab at the neck level (green in Fig. 1A). As the amount of labeled water is, in this case, dependent of the blood velocity, special care has to be taken in order to have the label bolus width proportional to subject- and state-specific flow. This is typically achieved by applying an additional RF pulse some time after the labeling, for example, in the QUIPSS II approach (red in Fig. 1A) (Luh et al., 1999), in order to saturate the water spins still present in the labeling slab. That is, for higher flow, more labeled spins exit the labeling slab before the application of the saturation RF pulse and, hence, the bolus duration is larger for higher flow. In CASL, RF is applied continuously in a plane also at the neck level, and all the spins following through that plane experience the labeling RF (Detre et al., 1992). This approach naturally takes care of the bolus duration and no additional RF pulse to cut off the bolus end is necessary. In pCASL, instead of a continuous RF irradiation, a train of RF pulses is applied, and it has been shown that this results in a higher labeling efficiency (Dai et al., 2008).

ASL approaches only effectively measure CBF, if imaging artifacts are avoided. The most important obstacles are: magnetization transfer (MT), arterial blood contribution due to incomplete delivery of labeled blood to the tissue and venous labeling.

MT effect due to off-resonance RF excitation and incomplete delivery of arterial blood to the tissue due to short post-labeling delay (PLD) may result in ASL MRI signal stemming from the arteries rather than capillaries or tissue and thus, spatially bias cortical-depth-dependent CBF-fMRI. Hence, they are usually accounted for by means of balanced off-resonant pulses (Pekar et al., 1996) and appropriate choice of PLD (Alsop et al., 2015), respectively. Unwanted venous blood labeling, such as often observed in the FAIR approach (Çavuşoǧlu et al., 2009; Wong et al., 1997), could be accounted for with additional crusher gradients. In high-resolution, cortical depth-dependent ASL fMRI, however, this is usually not applied because of possible reduction of ASL SNR, due to altering the amount of label present in the laminar-specific capillaries. Instead, at high resolutions, individual voxels of large venous vessels can be clearly detected and excluded from further consideration in cortical depth-dependent CBF studies.

Other sources affecting the ASL signal, such as differences between blood and tissue, inhomogeneities of the magnetic field and flow velocity variation at the tagging plane/slab, will result in erroneous absolute CBF estimates. However, the effect of these sources on the laminar profile (i.e. the cortical depth distribution of the relative CBF) have not been shown to be relevant in first approximation, as all the ASL signal of all layers are similarly affected by these sources. These error sources may become relevant if absolute CBF values are compared between brain areas. Future research has to quantitatively determine the influence of these and other sources on both the laminar CBF profiles and quantitative CBF estimates.

2.2.2. Early animal and human studies at high spatial resolution

Laminar fMRI has already been performed in animals using ASL. The group of Seong-Gi Kim performed the first study with laminar ASL in cats at 9.4T albeit with primary focus on arterial versus venous CBV changes (Kim et al., 2007). In another study, they demonstrated for different phases of the response (4–10s, 12–30s and 32–60s after stimulus start for 60s stimulation) GM localization of both CBF and CBV with peak in presumed layer IV of primary visual cortex (V1) (Jin and Kim, 2008a). Independent of the time period, the BOLD signal was maximal on the surface of the brain. Post-stimulus undershoot was observed both in the BOLD signal and CBF. Zappe et al. investigated the laminar distribution of CBF and compared it with that of the GE-BOLD signal in macaque monkeys at 4.7T and 7T (Zappe et al., 2008). In order to achieve high labeling efficiency, CASL with a separate labeling coil was employed. They also found the well-established increase in BOLD signal towards the surface of the brain. In contrast, functional CBF exhibited a peak in the middle cortical depths. Kim and colleagues found in the rat that CBF functional responses are spatially specific to well-known neuronal subdivisions in the olfactory bulb in contrast to the BOLD response (Poplawsky et al., 2015). These two studies strongly argue that laminar CBF and CBV imaging are feasible and specific to local neuronal populations.

In humans, to the best of our knowledge, only two studies using (sub)millimeter in-plane resolution (albeit with supra-millimeter slice thickness) have been performed, both published in 2002. Pfeuffer et al. showed the feasibility of high-resolution ASL with 0.9×0.9×1.5 mm³ FAIR at 7T (Pfeuffer et al., 2002). A dedicated half-volume coil and adiabatic RF pulse enabled homogeneous labeling with a reduced field-of-view (FOV) imaging to push the spatial resolution while keeping the read-out duration short. They demonstrated detectable CBF during both baseline and functional activation. As predicted, CBF had a better spatial localization with respect to tissue than the BOLD signal, which was most strongly activated in surface veins. Duong et al. also used FAIR and a reduced FOV with spatially selective, orthogonal RF pulses (Duong et al., 2002). These experiments were performed both at 4T and 7T. In-plane resolution of BOLD data was either 0.5×0.5 or 1×1 mm², and for CBF 1×1 mm² or 1.5×1.5 mm². Slice thickness was 1 or 2 mm for BOLD and 3 or 4 mm for CBF imaging, respectively. Robust and highly reproducible CBF activation was detected in the visual and motor cortex. Note, however, that both studies did not attempt to study the laminar profile of CBF activation.

The animal studies using ASL clearly demonstrated that CBF functional response is laminar specific with peak response in the middle cortical depths, which is expected to be the location of the highest neuronal activation following sensory stimulation. However, the exact spatiotemporal profile of CBF and its stimulus and brain-state dependency have not been studied in humans yet due to lack of availability of advanced ASL approaches at 7T and low signal-to-noise ratio (SNR) of these techniques.

2.2.3. Estimation of laminar SNR of ASL

At 1.5 and 3T for standard resolutions in humans (e.g., 3–5mm isotropic), it has been shown that pCASL has higher SNR compared to PASL and CASL (Alsop et al., 2015). However, for laminar imaging, it is desirable that ASL is performed at 7T or higher field to achieve submillimeter resolution. For baseline perfusion mapping, pCASL has been implemented at 7T (Ghariq et al., 2012; Luh et al., 2013; Wang et al., 2015). Due to the fact that the specific absorption rate (SAR) is higher for pCASL than for PASL, high-field pCASL implementations are currently limited to repetition times $T_R>5\mathrm{s}$. Hence, at 7T, functional experiments have so far only been conducted with PASL. We believe that with the emergence of routinely applicable parallel RF transmit approaches, more SAR-efficient pCASL setups might become available for future applications of high-res CBF-fMRI however.

For ideal conditions, using QUIPSS II, the signal difference due to labeling, which is proportional to CBF, is (Wong et al., 1998):

$$\frac{\mathrm{\Delta }{S}_{\mathrm{A}\mathrm{S}\mathrm{L}}}{S_{0,B}}=2\cdot f\cdot T_{I,1}\cdot e^{-\frac{T_{I,2}}{T_{1,B}}}\cdot e^{-\frac{T_E}{T_2^{\mathrm{*}}}},$$ (1)

where, $\mathrm{\Delta }{S}_{\text{ASL}}$ denotes the signal difference between control and label images, $S_{0,B}$ the (baseline) signal intensity of fully relaxed blood, $f=\mathrm{C}\mathrm{B}\mathrm{F}/\lambda$ the blood flow in units of ${\mathrm{s}}^{-1}$ (order of $0.01{\mathrm{s}}^{-1}),\lambda$ the blood-brain partition coefficient, $T_{I,1}$ the bolus duration (order of 0.8 s ), $T_{I,2}$ the PLD (order of 1.8 s ) and $T_{1,B}$ the $T_1$ value of blood, which is approximately 2.1s at 7T (Zhang et al., 2013). Imaging is typically done at $T_E=15\mathrm{m}\mathrm{s}$ and with $T_2^{\mathrm{*}}\approx 30\mathrm{m}\mathrm{s}$. Inserting these values into Eq. 1, we obtain $\mathrm{\Delta }{S}_{\mathrm{A}\mathrm{S}\mathrm{L}}/S_{0,B}\approx 0.41\mathrm{\%}$. Thus, if we assume for a high-contrast stimulus an average CBF change of 50% (e.g., (Uludag et al., 2004)), the activation-induced signal change in units of baseline signal is 0.6%.

We estimate the temporal SNR (tSNR) of CBF at 7T with respect to the tSNR of laminar GE-BOLD signal in order to determine the feasibility of laminar CBF imaging. In a recent study, we determined for 0.7mm nominal isotropic resolution for high-contrast checkerboard stimulation a BOLD signal of 2–4% $T_E=26\mathrm{m}\mathrm{s}$ for the middle GM voxels (Kashyap et al., 2017). The percent BOLD signal amplitude is, to first order, proportional to $T_E$ and is acquired at a longer $T_E$ than CBF acquired with ASL. Thus, the interpolated BOLD signal amplitude at $T_E=15\mathrm{m}\mathrm{s}$ would be 1–2%. Consequently, laminar CBF tSNR for imaging at 0.7mm isotropic resolution is expected to be lower by a factor of 5–10 than that for BOLD fMRI.

2.3. Results of CBF across cortical depth

In the following, we show illustrative examples that functional and baseline laminar CBF is feasible. Fig. 1 depicts perfusion maps (panel B) and depth-dependent profiles (panel C) acquired with a nominal isotropic resolution of 0.7 mm in the human visual cortex (Ivanov et al., 2016b). It can be seen that perfusion is highest inside the GM close to the middle cortical depths, suggesting that it is not dominated from macro-vessels outside the GM. Functional results depicted in Figs. 1D–E of one representative participant show that the highest CBF activation is located in deeper cortical depths as compared to the BOLD signal (green arrows). CBF values are smaller in upper cortical laminae, which suggests that CBF-fMRI is not biased by large pial vessels. The fact that CBF is mostly dominated from tissue vessels inside the GM is consistent with the previously shown cortical profiles of baseline CBF and of CBF changes in rats (see Fig. 4 in (Shen et al., 2008) and Fig. 6 in (Shen et al., 2015)), in cats (see Fig. 1 in (Jin and Kim, 2008a)), or in monkeys (see Fig. 6 (Goense et al., 2012a).

The amount of CBF change is approximately 40%. This refers to a control–label signal difference of approximately 1.5%, that is, the CBF change in units of baseline EPI signal is approximately 0.6%.

3. Cerebral blood volume

There are several different methods proposed and applied for CBV-fMRI: iron-oxide-CA-based, gadolinium-CA-based and non-invasive vascular space occupancy (VASO)-based methods. Despite a few pioneering functional studies (Belliveau et al., 1991; Frahm et al., 2008), gadolinium CAs have not yet found application for depth-dependent fMRI and will, thus, not be further considered here.

3.1. Iron-oxide-based CBV-fMRI

3.1.1. Principle of iron-oxide-fMRI

For depth-dependent CBV-fMRI, iron oxide is used as an intravascular susceptibility agent with long biological half-life. Iron-oxide CAs usually consist of particle sizes in the order of 4–10 nm (Shen et al., 1993) and they show superparamagnetic properties. Upon intravenous injection, they generate a strong $T_2^{\mathrm{*}}/T_2$ contrast, which alters the (water) signal in relation to CBV. Due to faster magnetization dephasing in the presence of intravascular iron oxide, the corresponding MRI signal reduction can be considered to primarily reflect the dynamics of CBV modulated by neural activity. These signal changes can be modeled analytically (Kiselev and Posse, 1999; Yablonskiy and Haacke, 1994) or with Monte Carlo simulations (Boxermann et al., 1995; Martindale et al., 2008). In first order approximations, relative iron-oxide-weighted signal decreases are inversely proportional to relative CBV increases. It must be noted that the iron-oxide-modulated MRI signal at baseline is already reduced proportional to the regional baseline blood volume, CBV₃. Hence, the signal change during neural activity only reflects relative CBV changes (in %), but not absolute volume changes (in ml).

There are currently no iron-oxides approved for MRI in human subjects in non-clinical protocols (Mandeville, 2012). However, ferumoxytol has been approved in the US for the treatment of iron deficiency anemia in adult patients with chronic kidney disease (Lu et al., 2010). Hence, within corresponding clinical protocols, it could be deployed for CBV-fMRI in control groups of healthy volunteers. The recent occurrence of adverse events (including 18 cases of death since 2009), however, has caused the FDA to approve a Boxed Warning, the FDA’s strongest kind of warning. It is explicitly stating that “Feraheme is specifically approved for use only in patients with iron deficiency anemia with chronic kidney disease”. (www.fda.gov/Drugs/DrugSafety/ucm440138.htm).

While the first human fMRI applications with iron-oxide CA were only aiming at routine spatial resolutions (D’Arceuil et al., 2013; Qiu et al., 2012), the promising potential of iron-oxide CA for high-resolution depth-dependent fMRI is still to be confirmed.

The fMRI acquisition parameter space of cortical-depth-dependent fMRI is very similar compared to conventional GE-BOLD-fMRI. Both temporal and spatial resolution are mostly limited by hardware, including the gradient speed and parallel imaging acceleration performance of the receive coil. The only parameter that should be adapted compared to BOLD fMRI is the slightly shorter optimal $T_E$.

Note that the application of fMRI with iron-oxide nanoparticles is known under several names, including UPSIO (Kim et al., 2013), IRON (Mandeville, 2012), ferumoxytol (Christen et al., 2012), or MION (Shen et al., 1993).

3.1.2. Previous depths-dependent iron-oxide-based CBV-fMRI

The first iron-oxide-based CAs have been developed in the late 80s (Weissleder, MRM 1992), and their application for fMRI in animals has been pioneered at MGH (Mandeville et al., 1998). The motivation to use them as a measure of neural activity changes came from the failure to obtain satisfactory BOLD results outside large pial vessels (Mandeville, 2012). Already in the very early studies, the cortical-depth dependence of CBV-fMRI was investigated and compared to that of BOLD-fMRI showing that the BOLD signal is more dominantly affecting the fMRI signal at the cortical surface compared to CBV (Mandeville et al., 2001, 1998; Mandeville and Marota, 1999).

Because of the higher contrast-to-noise ratio (CNR) of iron-oxide-fMRI compared to BOLD-fMRI, it quickly became the method of choice for laminar fMRI in animal models. Multiple research labs followed the initial experiments and started to deploy iron-oxide-fMRI in a more rigorous approach and widespread analysis of laminar CBV responses. For instance, laminar CBV-fMRI has become the focus of attention of groups around Kim (Kim et al., 2013) investigating cats and rats, Harel and Yacoub (Harel et al., 2006) investigating cats, Silva (Silva et al., 2007), Kennerley and Mayhew (Kennerley et al., 2005), and Lu (Lu et al., 2004) all investigating rats, and Logothetis (Goense et al., 2007; Smirnakis et al., 2007) investigating monkeys. Those studies started with comparisons of ‘peak voxel locations’ across cortical depth in CBV and BOLD-fMRI (Mandeville et al., 1998; Mandeville and Marota, 1999; Shen et al., 2008) and developed towards complete depth-dependent profiles of CBV-fMRI (Goense et al., 2007; Harel et al., 2006; Kennerley et al., 2005; Lu et al., 2004). Later studies continued with increasing levels of sophistication:

1. Depth-dependent profiles were investigated with respect to the temporal evolution of the CBV signal during and after stimulation (Goense et al., 2012b; Jin and Kim, 2008a; Silva et al., 2007; Yacoub et al., 2006).

2. More effort was also focused on the quantitative interpretation of CBV-weighted profiles across cortical depth. Analysis methods evolved from initial ‘CBV-weighted activity’ towards semi-quantitative measures of percent CBV changes normalized by the background signal (Jin and Kim, 2008a; Kennerley et al., 2005; Kim and Kim, 2011a, 2010; Zhao et al., 2006). Ultimately, absolute CBV changes ($\mathrm{\Delta }\mathrm{C}\mathrm{B}\mathrm{V}$ in ml per tissue volume) were analyzed (Huber et al., 2015; Kim et al., 2013; Kim and Kim, 2011b).

3. A few initial studies have already started to exploit depth-dependent CBV responses in neuroscientific applications inferring directional connectivity information, for instance, in rat olfactory bulb (Poplawsky and Kim, 2014), in rat barrel cortex (Boorman et al., 2010) and monkey visual cortex (Goense et al., 2012b).

3.2. VASO-fMRI

3.2.1. Principle of VASO-fMRI

VASO is non-invasive and, hence, the most widely applied CBV-sensitive fMRI method in humans (Lu et al., 2013, 2003). It takes advantage of the $T_1$ difference between blood and surrounding tissue using an inversion-recovery sequence to null blood signal while maintaining part of the tissue signal. The (relative) VASO signal intensity can, thus, be considered to be proportional to 1–CBV (if no BOLD contaminations are present). When neural activation causes CBV to increase, the VASO signal will show a decrease, allowing the detection of activated regions. A schematic of the MRI sequence and the expected magnetizations are depicted in Fig. 2A.

Figure 2: CBV estimation with VASO.

(A) illustrates the contrast-generation mechanisms of VASO. Panels (B-D) depict fMRI results in the hand representation of M1 and were acquired at a nominal resolution of 0.75×0.75×1.5 mm³ with 3D-EPI VASO during a 12min finger-tapping (pitch motion of thumb and index finger) paradigm (30s on vs. 30s off) in one representative human participant at 7T. Panel (B) compares cortical-depth-dependent activity of VASO and BOLD signal. VASO-CBV results show indications of most activity in upper and secondary activity in deeper cortical laminae, while GE-BOLD is more dominated from upper cortical depths. Panel (C) depicts the characteristic temporal features of depth-dependent CBV-fMRI. With a temporal resolution of a few seconds, deeper laminae show a slower response. Panel (D) shows the same VASO data evaluated in different units.

Dependent on field strength and the interaction of thermal and physiological noise, VASO CNR is about 40%-70% of that of GE-BOLD-fMRI (Huber et al., 2014b; Lu et al., 2013). The exact value further depends on the contribution of physiological noise, resolution and tissue compartment. The smaller sensitivity of VASO compared to BOLD is often considered to be the major limitation for submillimeter fMRI studies.

Compared to the BOLD signal, the temporal resolution of VASO is inherently limited by its contrast-generation mechanism. Since CBV-weighting is obtained by means of $T_1$-dependent blood-signal nulling, the timescale of blood $T_1$ (1.6s at 3T, 2.1s at 7T, 2.4s at 9.4T) can be considered as an approximate limit of the minimal repetition time $T_R$. Furthermore, contrast mechanisms in VASO modifications that are specifically optimized for high fields (Huber et al., 2014b) additionally dictate that $T_R$ be longer than the vascular refill time. Hence, temporal resolutions well below 3s are scarce in the VASO literature.

Due to the required blood-nulling condition, the functional signal must be captured during a relatively short acquisition window. If the acquisition duration of individual slices would be too long, they would end up with different effective inversion times ($T_I$) with insufficient blood-signal nulling, limiting the brain coverage of the method. This is especially problematic for high-resolution protocols with long acquisition trains. Hence, past cortical-depth-dependent VASO-fMRI has been based on advanced readout strategies to partly account for this limitation, including 3D-EPI and SMS (Huber et al., 2016c).

With increasing field strength as often used for submillimeter resolution, the positive BOLD signal change during neural activation increasingly counteracts the negative VASO signal change (Lu and van Zijl, 2005). This BOLD contamination is heterogeneous across cortical depth and, thus, can introduce an unwanted bias in the interpretation of laminar VASO results. Hence, BOLD contaminations must be accounted for. In past high-resolution studies this has been done with two approaches: (A) By minimizing $T_E$ with multi-shot EPI (Donahue et al., 2006) and center-out k-space trajectories (Jin and Kim, 2008b). (B) By concurrent interleaved acquisitions of BOLD and VASO signals (Huber et al., 2015) or multi-echo EPI (Huber et al., 2014a).

3.2.2. Previous laminar VASO-fMRI in humans

Initial high-resolution CBV data in humans were acquired with 0.78×0.78×3 mm³ voxels (Donahue et al., 2006). Comparisons of depth-dependent VASO and iron-oxide profiles in cat striate cortex showed mostly corresponding results (Jin and Kim, 2008b), with residual discrepancies in upper cortical laminae. Later, quantitative comparisons in rat sensory and monkey visual cortex resolved this inconsistency to be a result of the different dependence on baseline CBV in iron-oxide and VASO-fMRI (Huber et al., 2015). That is, depth-dependent CBV profiles are identical for VASO and iron-oxide fMRI if expressed in absolute units. The first depth-dependent CBV-fMRI analyses in humans were limited to two laminae—an upper lamina with versus a deeper lamina without pial vessels (Huber et al., 2014a, 2014b). More recent advancements in fMRI readout strategies facilitates the measurement of cortical profiles with up to four (Huber et al., 2015) and seven voxels (Huber et al., 2016c) across the depth of primary motor cortex (M1). This resolution made it possible to address questions about depths-dependent feed-forward versus feedback connectivity in cases of neural inhibition (Huber et al., 2015, 2014a) and efferent- versus afferent-driven activity during modulated sensory-motor tasks and at rest (Huber et al., 2016a).

3.3. Expected microvascular origin of laminar CBV-fMRI

Recent high-resolution optical studies have established the following picture with regards to CBV changes: Pial arterioles above the cortical surface can increase their volume by up to 30% providing a significant (Drew et al., 2011) or sometimes even dominant CBV contribution (Kennerley et al., 2012). The smaller, depth-unspecific diving arterioles provide another significant contribution to the overall CBV change (Hillman et al., 2007; O’Herron et al., 2016; Tian et al., 2010). Most of the CBV change, however, is expected to come from small arterioles that are located close to the layer-dependent neurons (see Fig. 4 below). Among those, the relative contribution of CBV change slightly increases with higher branching order and smaller vessel diameter (Drew et al., 2011; Gagnon et al., 2015). Further downstream compartments, in particular pial veins, are believed to have limited influence on the overall CBV change (Drew et al., 2011; Gagnon et al., 2015; O’Herron et al., 2016; Takano et al., 2006). However, there is currently not much data available on how much venules, located in the microvasculature, and ascending veins change their CBV following functional stimulation. Short stimuli were shown to result in negligible pial venous dilation (Hillman et al., 2007; Takano et al., 2006). However, some studies report that longer stimuli can evoke some pial venous CBV increase below that from the arterial components (Drew et al., 2011). It is to note that the terminology of ‘arterial’, ‘arterioles’, ‘capillary’, ‘venules’ and ‘venous’ CBV is not consistently used across research groups and imaging modalities. Microscopy studies can distinguish arteries, capillaries, and veins based on the cytoarchitecture of the vessel walls. FMRI and optical imaging spectroscopy studies, however, use terminologies that refer to the blood oxygenation rather than the cytoarchitecture. In most fMRI and optical spectroscopy studies, fully or almost fully oxygenated blood (oxygen saturation 90–100%) is considered to be ‘arterial’, while partly oxygenated blood (oxygen saturation 50–70%) is considered to be more ‘venous’ and values in between are attributed to capillaries.

Figure 4: Schematic comparison of depth-dependent ASL, VASO, calibrated fMRI and GE-BOLD-fMRI.

The left column illustrates the vascular origin across large vessels and small, depth-specific microvessels for BOLD (A), ASL (B), VASO (C), and calibrated fMRI (D). For GE-BOLD-fMRI, most of the oxygenation and signal changes are believed to originate from large, depth-unspecific veins (A). For ASL, the functional signal of the labeled blood is believed to be dominated from the depth-specific permeable capillary vessels (B). For VASO, the signal change is believed to result from all vaso-active vessels including depth-unspecific arterioles and depth-specific arterioles and capillaries (C). For calibrated fMRI, the grey background suggests that ${\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$ estimates are expected to arise from extravascular tissue properties only (D). The right column depicts a summary of the most important quality features of BOLD, ASL, VASO and calibrated fMRI. The presented values are based on the summary of Tab. 1 and reflect the detailed discussion in section 5. All six compared parameters are normalized to the value of the fMRI method yielding the highest value.

3.4. Spatiotemporal features of depth-dependent CBV-fMRI

Fig. 2 depicts all the most prominent characteristics of laminar CBV: cortical depth-dependent profiles, temporal features across depth, signal and noise distributions, as well as specific features of CBV changes when evaluated in different physical units.

3.4.1. Depth-dependence of CBV responses compared to GE-BOLD-fMRI

Representative profiles of CBV and BOLD signal changes are shown in Fig. 2B. While the dominant BOLD contrast is obtained at upper cortical depth and above the cortical surface, $\mathrm{\Delta }\mathrm{C}\mathrm{B}\mathrm{V}$ has a strong functional contrast deeper inside the cortex as well. The fact that CBV does not have the strong sensitivity to large superficial vessels is established in the field and similar GE-BOLD versus CBV profile comparisons are abundantly shown in the literature (Donahue et al., 2016; Huber et al., 2015, 2014a; Kennerley et al., 2005; Kim et al., 2013).

3.4.2. Different timing of hemodynamic responses across cortical depth

Fig. 2C indicates that the upper cortical laminae have a faster CBV response compared to deeper ones. This suggests that cortical profiles might look slightly different for different stimulation durations due to different weighting of larger and smaller arterial and capillary vessels across cortical depth. Similar depth-dependent behavior was reported from animal studies (Huber et al., 2016b, 2014a; Kim et al., 2013; Kim and Kim, 2011b; Silva et al., 2007; Yacoub et al., 2006).

3.4.3. Profiles of CBV change evaluated in different units

In Fig. 2D, the effect of heterogeneous baseline in depth-dependent fMRI signal is visualized by means of showing the identical results in different units. Obviously, when evaluated in absolute units (ml per 100 ml of tissue), $\mathrm{\Delta }\mathrm{C}\mathrm{B}\mathrm{V}$ can be highest for upper cortical depths whereas percent units ($\mathrm{\Delta }\mathrm{C}\mathrm{B}\mathrm{V}/{\mathrm{C}\mathrm{B}\mathrm{V}}_0$) result in low values for upper peak positions relative to the deeper peak. This is due to the larger denominator of ${\mathrm{C}\mathrm{B}\mathrm{V}}_0$ in upper cortical laminae.

The approximate change in CBV in the range of 3.5 ml/100ml, refers to a 3.5% change in the VASO time course signal.

3.5. Confounding factors in depth-dependent VASO-fMRI

3.5.1. Perfusion contaminations

Due to the finite permeability of capillary walls, the VASO two-compartment model of blood and tissue (Lu et al., 2003) is not accurately reflecting typical brain physiology, and downstream intravascular space might contain non-nulled tissue water magnetization. The resulting magnitude of this effect has been quantified in simulations to be negligibly small, below 3.5% (relative to the relative VASO signal change in %) (Wu et al., 2010). This means that a typical VASO signal decrease during activation of −1.5% can only be assessed with an accuracy of (−1.500 ± 0.053)%.

3.5.2. Blood T₁-uncertainties

It has been recently shown that arterial blood $T_1$ is approximately 100–200 ms longer than venous blood $T_1$ (Grgac et al., 2012; Rane and Gore, 2013). Consequently, the chosen $T_I$ can differ from the true blood-nulling time by 35–70 ms. The correspondingly incomplete blood-signal nulling can result in an error in the absolute VASO signal of the tolerable amount of 0.05% (see discussion in (Huber et al., 2016b; Wu et al., 2010)). This suggests that a typical VASO signal decrease in the order of −1.5% can only be assessed with an accuracy of (−1.50 ± 0.05)%.

3.5.3. CSF-contaminations

Functional changes in the partial voluming of cerebrospinal fluid (CSF) and GM can affect the VASO signal (Donahue et al., 2006; Jin and Kim, 2010; Piechnik et al., 2009) and can mimic a VASO signal change at the cortical surface and distort the corresponding cortical profiles. This potential contamination can be accounted for by manipulations of relative $T_1$ weighting of CSF, blood and GM (Donahue et al., 2006; Huber et al., 2016b; Scouten and Constable, 2008).

4. Cerebral metabolic rate of oxygen

Although the brain accounts for only 2% of the human body mass, it consumes approximately 20% of the energy under awake, resting conditions (Mink et al., 1981). This demand is predominantly covered by oxidative metabolism of glucose, which is closely coupled to neurotransmitter cycling and neuronal firing over a wide range of cortical activity (Shulman et al., 2001; Sibson et al., 1998). Hence, the investigation of brain energetics expressed as ${\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$ provides a physiologically well-defined, quantitative link to neuronal activity (Lin et al., 2010; Shulman and Rothman, 2005). Moreover, microelectrode studies demonstrated that tissue oxygenation changes occur on a spatial scale of single-unit activity (Thompson et al., 2005), which suggests that ${\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$ mapping has also a potential for imaging brain activity with high spatial specificity.

4.1. Mapping of oxygen and glucose utilization with tracer techniques

Spatially resolved investigations of cerebral energetics became possible with the development of 2-[¹⁴C]deoxyglucose (2DG) autoradiography (Kennedy et al., 1974; Sokoloff et al., 1977), allowing to map the cerebral metabolic rate of glucose, ${\mathrm{C}\mathrm{M}\mathrm{R}}_{\text{GIc}}$, which is coupled to ${\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$, at a resolution in the order of 100 μm (Sokoloff, 1981). Early work achieved delineation of cortical columns in macaque V1 (Kennedy et al., 1976) or rat primary somatosensory cortex (S1) (Kossut et al., 1988) and showed highest ${\mathrm{C}\mathrm{M}\mathrm{R}}_{\text{GIc}}$ in layer IV, which is not particularly rich in neuronal cell bodies but in synaptic connections. This suggests that it is mainly the energetically expensive synaptic activity that is linked to activation-induced glucose consumption (Attwell and Laughlin, 2001). Because synaptic density and electrical and chemical activities vary across cortical layers (Hyder et al., 2013), we may expect a related cortical-depth-dependent signature of ${\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$.

An analogue to the 2DG technique for use in human subjects in vivo was obtained with the development of positron emission tomography (PET) (Phelps et al., 1979; Reivich et al., 1977). With estimated resolution limits of 0.83 mm and 2.36 mm for preclinical and clinical scanners, respectively (Moses, 2011), PET has a potential for studying laminar activation in small animals but not in human subjects.

Results from radiotracer studies on brain energetics have been confirmed and expanded in vivo by ¹³C MR spectroscopy by following the flux of ¹³C isotope into cerebral metabolites during intravenous infusion of ¹³C-labeled glucose or other substrates (Sibson et al., 1997). This has been used in animal (Hyder et al., 1997; Mason et al., 1992; Sibson et al., 1998) and human brain (Gruetter et al., 2001; Mason et al., 1999, 1995; Shen et al., 1999) to measure the rate of the tricarboxylic acid cycle. The results show that approximately 80% of oxidative glucose consumption is related to signaling at baseline and an even higher percentage during activation (Hyder et al., 2013; Hyder and Rothman, 2012; Rothman et al., 1999; Shulman et al., 2001). However, these experiments require long scan times (1 h or more) due to relatively slow turnover rates and are limited in their spatial resolution (order of millimeters in rats or centimeters in humans) due to the inherently low sensitivity.

4.2. Calibrated fMRI

4.2.1. Relative measurements of changes in oxygen consumption

The earliest and most widely employed approach to extract information on ${\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$ from the BOLD response relies on additional ASL measurements of CBF and the application of a physiological model (Davis et al., 1998; Hoge et al., 1999a, 1999b; Kim et al., 1999). In the Davis model (Davis et al., 1998), employing Fick’s principle, the normalized GE-BOLD signal change upon activation, $\mathrm{\Delta }{S}_{\mathrm{B}\mathrm{O}\mathrm{L}\mathrm{D}}$, is approximated as:

$$\frac{\mathrm{\Delta }{S}_{\mathrm{B}\mathrm{O}\mathrm{L}\mathrm{D}}}{S_0}=M\left[1-\left(\frac{{\mathrm{C}\mathrm{B}\mathrm{V}}_v}{{\mathrm{C}\mathrm{B}\mathrm{V}}_{v,0}}\right){\left(\frac{\mathrm{C}\mathrm{B}\mathrm{F}}{{\mathrm{C}\mathrm{B}\mathrm{F}}_0}\right)}^{-\beta }{\left(\frac{{\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}}{{\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{02,0}}}\right)}^{\beta }\right],$$ (2)

where subscripts ‘$0$’ and ‘$v$’ denote the baseline state and the venous compartment, respectively, and $\beta$ accounts for the nonlinear dependence of the field-inhomogeneity-induced transverse relaxation rate, $R_2^{\prime }$, on the deoxyhemoglobin concentration, [dHb], (Boxermann et al., 1995; Ogawa et al., 1993; Yablonskiy and Haacke, 1994).$M$. represents the maximum BOLD signal change that would be obtained with a venous oxygen saturation of 100% reflecting the inherent ‘ceiling’ of the BOLD response (Buxton, 2010). It depends on a plethora of physical, anatomical and physiological parameters, including the magnetic field strength, $T_E$, vessel geometry, susceptibility difference between blood and tissue, ${\mathrm{C}\mathrm{B}\mathrm{V}}_0$ and $[\mathrm{d}\mathrm{H}\mathrm{b}{]}_0$. Hence, it can vary between subjects or brain areas. As a further simplification, it is often assumed that CBV and CBF are coupled according to an empirical power law (Grubb et al., 1974):

$$\mathrm{C}\mathrm{B}\mathrm{V}\propto {\mathrm{C}\mathrm{B}\mathrm{F}}^{\alpha },$$ (3)

which allows substituting CBF for CBV in Eq. 2. A Grubb’s constant of ${\alpha }_t=0.38$ has been measured with PET in macaque brain for the relation between total CBV and CBF (Grubb et al., 1974). However, the largest changes in [dHb] upon activation occur in the venous compartment, where changes in CBV are relatively small (Hillman et al., 2007; Huber et al., 2014a; Lee et al., 2001). Hence, use of ${\alpha }_t=0.38$ is likely to underestimate $\mathrm{\Delta }{\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$ as evidenced by simultaneous recordings of BOLD, CBF and CBV responses (Chen and Pike, 2010a, 2009; Lin et al., 2008). This has been addressed by using a smaller constant, ${\alpha }_v\approx 0.2$, to model the relation between venous CBV and CBF (Chen and Pike, 2010a, 2009):

$$\frac{\mathrm{\Delta }{S}_{\mathrm{B}\mathrm{O}\mathrm{L}\mathrm{D}}}{S_0}=M\left[1-{\left(\frac{\mathrm{C}\mathrm{B}\mathrm{F}}{{\mathrm{C}\mathrm{B}\mathrm{F}}_0}\right)}^{{\alpha }_v-\beta }{\left(\frac{{\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}}{{\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}\mathrm{2,0}}}\right)}^{\beta }\right].$$ (4)

In calibrated BOLD experiments, $M$ is measured employing reference scans that involve episodes of hypercapnia and/or hyperoxia (Blockley et al., 2013; Hoge, 2012; Pike, 2012). The corresponding experimental setup is schematically depicted in Fig. 3A. A typical calibration consists of the administration of a gas mixture with elevated carbon dioxide (e.g., 5% CO₂ and 21% O₂ balanced with nitrogen) without stimulus presentation (Davis et al., 1998). It is assumed to induce a purely vascular response but minimal effects on oxygen consumption (Chen and Pike, 2010b), whereby the last term drops out of Eq. 4. $M$ is, thus, obtained from simultaneous recordings of the BOLD and CBF responses in the calibration scan and inserted into Eq. 4 to yield an estimate of (relative) ${\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$ during the stimulation. Alternatively, VASO measurements to substitute CBV for CBF have also been employed (Guidi et al., 2016b), which yields:

$$\frac{\mathrm{\Delta }\mathrm{S}}{S_0}=M\left[1-{\left(\frac{{\mathrm{C}\mathrm{B}\mathrm{V}}_t}{{\mathrm{C}\mathrm{B}\mathrm{V}}_{t,0}}\right)}^{\frac{{\alpha }_v-\beta }{{\alpha }_t}}{\left(\frac{{\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}}{{\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}\mathrm{2,0}}}\right)}^{\beta }\right].$$ (6)

Note that the Davis model was not developed for submillimeter resolution and, in addition, neglects intravascular contributions to the BOLD signal, which is reasonably well fulfilled at 7T but not at lower field (Boxermann et al., 1995; Jochimsen et al., 2004).

Figure 3: CMR_O2 estimation with calibrated fMRI.

(A) illustrates the working principle. Panels (B-D) depict profiles from a 15min unilateral five-finger-tapping task averaged across 11 participants with a nominal in-plane resolution of 0.8 mm at 7T. As seen in Fig. 3, VASO has a strong response in upper cortical laminae (presumably layers III) inside GM and a shoulder of CBV change in deeper cortical laminae before it decreases towards WM. The BOLD response, on the other hand, is dominated from the pial veins close to the cortical surface. The profile of (percent) ${\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$ change across cortical depths is depicted in green with biggest functional change in upper-to-middle cortical laminae.

As an alternative to hypercapnia, hyperoxic calibration seeks to change venous oxygenation without affecting CBF and ${\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$ (Chiarelli et al., 2007a). Inhalation of O₂-enriched air increases the amount of oxygen dissolved in arterial plasma with a concomitant increase in venous oxyhemoglobin. This calibration requires additional recordings of end-tidal oxygen partial pressure to model the [dHb] change and, typically, CBF measurements to account for a small reduction in flow caused by vasoconstriction (Bulte et al., 2007a, 2007b). Finally, the BOLD response might be calibrated without respiratory manipulations drawing upon biophysical modeling and a measurement of $R_2^{\prime }$ (Blockley et al., 2015a, 2012). A thorough comparison of these calibration methods indicated that the hypercapnia method is most robust against confounding effects from venous CBV changes with activation (Blockley et al., 2012).

Finally, it has been suggested to use the information from so-called resting-state (rs) fMRI recordings for normalizing the BOLD response (Kannurpatti and Biswal, 2008). While there is currently no consensus about the best way how to extract a scaling factor from rs-fMRI scans and more thorough validation is needed, initial laminar-dependent investigations demonstrated high correlations of cortical profiles of the temporal standard deviation of the BOLD-response to hypercapnia and of $M$ with the fluctuation amplitude of rs-fMRI time series (Guidi et al., 2016a).

4.2.2. Mapping of baseline oxygen consumption

Calibration experiments discussed so far yield only relative ${\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$ changes evoked by a stimulus, whereas absolute measurements require additional knowledge of baseline oxygen consumption. This is important to keep in mind when it comes to interpretation of laminar ${\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$ results. Since the baseline ${\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$ might be heterogeneously distributed across cortical depth, cortical profiles of absolute ${\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$ can look different compared to profiles of relative ${\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$.

Gauthier and Hoge (Gauthier and Hoge, 2013, 2012) proposed a generalized calibration model linking the BOLD response, CBF, and end-tidal O₂ for arbitrary combinations of hypercapnia and hyperoxia to image $M$ and ${\mathrm{O}\mathrm{E}\mathrm{F}}_0$, which, after multiplication by ${\mathrm{C}\mathrm{B}\mathrm{F}}_0$, yields ${\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}\mathrm{2,0}}$. An alternative approach was proposed in (Bulte et al., 2012). Refinements include the integration of vessel-size information (Germuska and Bulte, 2014), which can be extracted from the same gas-manipulation experiments (Germuska et al., 2013; Jochimsen et al., 2010; Jochimsen and Möller, 2008).

Apart from BOLD-based techniques, quantitative susceptibility mapping can be used to image oxygen extraction along venous vessels utilizing the signal phase (Fan et al., 2014; Haacke et al., 2010; Xu et al., 2014). For additional discussion, we refer to (Blockley et al., 2015b; Fan et al., 2016).

4.3. Depth-dependent results of CMR_O2 obtained with calibrated fMRI

Changes in ${\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$ across cortical laminae evoked by electrical forepaw stimulation have been recently measured in rat S1 and compared to neural activity recorded simultaneously with microelectrodes (Herman et al., 2013). As expected, the BOLD response decreased monotonically from superficial to middle to deep lamina (all approximately 600μm thick and proposed to correspond roughly to anatomical layers I-III, IV-V, and VI, respectively). Similar to earlier observation (Shen et al., 2008), the opposite pattern was found for ${\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$ with the smallest value in superficial and a more subtle difference between middle and deep laminae. A corresponding spatial signature was observed for multiunit activity (MUA), which is believed to reflect output spiking activity, whereas CBF was stable across laminae with a pattern that was more similar to local field potentials (LFP). An inconsistent result in (Herman et al., 2013) compared to other depth-dependent studies was, however, the CBV profile that resembled that of the BOLD response.

Average laminar profiles of CBV, GE-BOLD and resulting ${\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$ changes recorded in human M1 during finger tapping (Guidi et al., 2016b) are shown in Fig. 3B–C. The maximum ${\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$ increase of roughly 35% (compared to its baseline) is located well inside GM. It is obtained from BOLD and VASO time series signal changes in the order of 3.5% and 5%, respectively. While an assignment to specific cortical layers remains speculative, a projection of the histologic lamination pattern typically found in human M1 onto the spatial axis suggests that the peak is found roughly within layer III. It is interesting to note that a consistent accentuation of glucose utilization in layers III-IV was obtained in rat motor cortex with 2DG autoradiography (Collins, 1978). Clearly, more work is needed to confirm these preliminary suggestions.

4.4. Confounding factors of depth-dependent calibrated fMRI

General error sources in calibrated fMRI have been extensively discussed and quantitatively estimated in (Blockley et al., 2013; Griffeth and Buxton, 2011; Gagnon et al., 2016; Guidi et al., 2016b) and include (A) nonlinear noise propagation in the exponential Davis model with low-tSNR input data; (B) a limited validity of Grubb’s CBF-CBV power-law relationship, as 5 was found to change dynamically during stimulation (Kida et al., 2007) and with cortical depth (Jin and Kim, 2008a); (C) $T_1$ changes caused by elevated plasma O₂ during hyperoxia; (D) potential ${\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$ variations due to deviations from an assumed isometabolic response to hypercapnia; (E) errors in assumed CBV₃ if ASL is substituted by VASO. These limitations are believed to be confined to a few percent of uncertainty without introducing specific bias across cortical depth. Simulations further indicate that the accuracy of the simple Davis model might be improved if $\alpha$ and $\beta$ are treated as adjustable parameters that capture several physical effects (Gagnon et al., 2016; Griffeth and Buxton, 2011; Wise et al., 2013).

4.4.1. Limited validity of the Davis model across cortical depth

Fick’s principle (and, hence, the Davis model) assumes that all oxygen delivered by CBF is either metabolized or drained away, $\left[\mathrm{d}\mathrm{H}\mathrm{b}\right]\propto {\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}/\mathrm{C}\mathrm{B}\mathrm{F}$. However, this does not properly consider that the individual terms are usually measured in different compartments of the vascular tree. That is, changes in [dHb] obtained from the BOLD signal predominantly arise from draining veins, whereas CBF or CBV estimates obtained with ASL or VASO, respectively, arise mainly from arterioles and capillaries. While the application of Fick’s principle is justified at low spatial resolution, where all vascular compartments are captured in the same voxel, it must be revisited at submillimeter resolution. For example, at the cortical surface, voxels are dominated by pial veins, and large BOLD responses may occur upon changes in [dHb] without concomitant changes observed by ASL or VASO. Consequently, physiologically unreasonable ${\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$ estimates can result in upper cortical laminae (Bohraus et al., 2011). In addition, substituting CBF with CBV or vice versa in the Davis model assumes a steady-state, which is difficult to experimentally determine as the stimulus-evoked dilation of the various vascular compartment follows different time constants (Goense et al., 2016; Krieger et al., 2012). Such effects need to be further investigated and are subject of current research.

5. Comparison of non-BOLD and standard GE-BOLD-fMRI

There are many criteria that need to be addressed in depth-dependent fMRI for reliable interpretation of laminar results. Below, GE-BOLD-, ASL-, VASO- and calibrated fMRI are compared with respect to the most vital quality features for depth-dependent applications in humans: (1) sensitivity, (2) specificity, (3) quantifiability, (4) sampling efficiency, (5) sequence availability and (6) invasiveness. This discussion is summarized in Fig. 4 and Table 1.

Table 1.

Comparison of quality attributes of ASL, VASO and calibrated fMRI compared to gold standard of GE-BOLD-fMRI.

Quality attributes of fMRI contrasts	GE-BOLD	ASL (CBF)	VASO (CBV)	Calibrated fMRI (CMRO2)
sensitivity	100% (gold standard)	10%-20% compared to BOLD	40%-70% compared to BOLD	5%-10% of that of BOLD
specificity	mostly large veins and limited contribution from small post-arterial vessels	permeable capillaries only	mostly small arterioles and limited contribution from large arteries	as specific as model and ASL/VASO data used
availability	virtually every scanner	most research scanners	6–10 research labs	gas calibration in 3–7 research labs
confounding factors	-	finite capillary, permeability, venous inflow, MT	accuracy of volume redistribution mechanism, CSF contaminations, $T_1$ uncertainties	Fick’s principle applicability at submillimeter resolution, uncertainties of Grubb‘s coefficient
best theoretical temporal resolution	1 s (limited by slew rate and acceleration performance)	2 s (label) and 2s (control)	1.5 s (blood nulling)	approx. 3–8 s (when functional steady-state is reached)
biggest advantage	highest sensitivity	capillary specificity absolute and relative changes detectable	good sensitivity-specificity compromise	Least dependent on neurovascular coupling and plumbing artefacts
invasiveness	MRI scanner environment: acoustic noise, no motion allowed, vertigo	MRI scanner environment	MRI scanner environment	MRI scanner environment, unpleasant gas inhalation

5.1. Sensitivity at high resolution

Sensitivity is one of the most limiting factors of cortical-depth-dependent non-BOLD-fMRI. ASL sensitivity is mostly limited by the relatively small effect size of the control–label signal difference resulting in about 10–20% of the tSNR achieved with BOLD-fMRI (see section 2). With VASO, signal stability is partly limited by GM signal attenuations as a result of the blood-nulling inversion pulse. The relatively small effect size of the functional signal change at locations of large vessels, however, results in a tSNR of about 40–70% compared to BOLD-fMRI. Calibrated fMRI inherits the sensitivity limitations of ASL or VASO and is additionally penalized by non-linear propagation of error in the exponential Davis model and the need to combine multiple experiments (calibration and neural task) to estimate ${\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$. The resulting tSNR is about 5%-10% compared to that of BOLD-fMRI.

With respect to the sensitivity limitation, we would also like to stress that for simultaneous measurements of BOLD signal and CBF, or BOLD signal and CBV, the higher sensitivity in the BOLD signal can be used to boost the sensitivity of the alternative contrast. For instance, Simon and colleagues have proposed a method to use the BOLD signal to improve the estimation of CBF fluctuations, in effect improving the SNR of CBF to be the same as of the BOLD signal but retaining the quantitative meaning of CBF in absolute units (Simon et al., 2013). However, this method is still in development and will have to deal with the challenges of localization of the different responses for laminar studies.

Note that, here, the quality feature of sensitivity is approximated as fMRI signal change with respect to the temporal signal fluctuations. It must be stressed, however, that any fMRI methods sensitivity should always be considered in combination with the complementary quality feature of specificity (as done below). For instance, if the sensitivity is driven by undesired unspecific promiscuous signal from large vessels, it might make more sense to use methods with lower sensitivity, as long as the lower signal refers to the desired locally specific microvasculature (Huber et al., 2017b). As such, Fig. 5b (Uludag et al., 2009) shows that at 7T, the GE-BOLD signal from macrovessels, including the ascending cortical veins, is approximately 10 times larger than the signal from microvessels. This suggests that the desired “microvascular sensitivity” could be actually higher in $\mathrm{C}\mathrm{B}\mathrm{F}/\mathrm{C}\mathrm{B}\mathrm{V}/{\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$ compared to the BOLD signal (e.g. due to ascending veins).

Sensitivity limitations with relatively weak more ‘cognitive’ tasks in future studies of non-BOLD fMRI might particularly benefit from extensive averaging across multiple sessions and multiple participants. To achieve this, advanced evaluation methods should be applied (Polimeni et al., 2017), including distortion matched anatomical reference acquisition (Huber et al., 2016d; Kashyap et al., 2017; Renvall et al., 2016) and participant-specific ‘laminar’ and ‘columnar’ coordinate systems (Huber et al., 2017a; Kemper et al., 2017).

5.2. Depth-dependent specificity

Below, we discuss specificity limitations of the individual contrast mechanisms. Additionally, any depth-dependent fMRI method is ultimately limited by specificity losses resulting from blurring during the signal readout (Kemper et al., 2016) and data processing (Kashyap et al., 2017), which are not in the scope of this review.

It is not straightforward to estimate specificity based on experimental non-BOLD fMRI results, because there is no established reference of depth-dependent signatures that other contrasts can be compared to. Hence, we discuss the specificity of the different fMRI contrasts based on the specificity of the vascular compartments they are most sensitive to. GE-BOLD signal is dominated from ascending and pial veins without depth signature and is, hence, often considered to be less specific than non-BOLD-fMRI methods (Uludaǧ and Blinder, 2017). Since functional signal in ASL arises from cortical-depth-specific permeable capillaries, it is believed to be as locally specific as the neurovascular coupling itself, with an estimated spatial scale of less than 0.2 mm in rats (Vazquez et al., 2014). VASO, however, is not only sensitive to depth-specific microvessels but also, to some extent, to unspecific diving and pial arteries. Hence, it might be considered to be less depth-specific than ASL, but more specific than GE-BOLD. ${\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$ is desired to be independent of neurovascular coupling and is, hence, presumably the most specific non-BOLD parameter. When determined by calibrated fMRI, however, it might be indirectly limited by the specificity of its underlying fMRI contrasts.

With respect to the smaller specificity and higher sensitivity of GE-BOLD compared to $\mathrm{C}\mathrm{B}\mathrm{F}/\mathrm{C}\mathrm{B}\mathrm{V}/{\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$, non-BOLD methods may be more comparable to SE-BOLD (see (Huber et al., 2017b) and other articles of this Special Issue).

In addition, any vascular fMRI contrast depends on the local coupling between neural activity changes, energy demand and vascular changes. Most of the fMRI-relevant energy demand refers to post-synaptic (Attwell and Iadecola, 2002) input-driven LFP changes (Logothetis et al., 2001). The neuronal cell bodies, however, can be located at different cortical depth compared to most of their dendrites (Cajal, 1906). Hence, the specificity of cortical depth-depended-fMRI relates to the layer specificity of dendrites compared to their somata.

5.3. Quantifiability of depth-dependent activity

The direct dependence of the fMRI signal on activity changes is often considered to be hampered by two separate information convolution mechanisms: (A) The physiological dependence of the vascular response on neural activity changes (i.e., neurovascular coupling). (B) The link of vascular responses to the observed fMRI signal. Since the susceptibility-sensitive BOLD signal is highly dependent on the exact interaction of multiple counteracting biophysical properties (e.g., CBF, CBV, ${\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$), its depth-dependent response is biased by both convolution mechanisms. ASL and VASO, however, provide quantitative estimates of the vascular response in biophysical units. They are limited by type-$A$ convolution but not type-$B$. ASL can be considered being even more quantitative than VASO, as it can straightforwardly quantify baseline blood flow in addition to $△\mathrm{C}\mathrm{B}\mathrm{F}$. The ${\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$ change estimated with calibrated fMRI is ideally independent of the neurovascular coupling and, hence, independent of type-$A$ and type-B convolutions. However, being based on ASL or VASO acquisitions, its quantifiability might be indirectly limited by these methods.

Another source of potential shortcomings of high-resolution BOLD methods compared to high-resolution non-BOLD contrasts is the fact that, compared to CBF/CBV/${\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$, the BOLD signal amplitude can be dependent on the orientation of the cortex with respect to main magnetic field (Fracasso et al., 2017; Gagnon et al., 2015).

5.4. Sampling efficiency

While the minimum $T_R$ of conventional BOLD-fMRI is mostly limited by gradient performance and achievable acceleration with parallel imaging, ASL, VASO and calibrated fMRI share additional limitations in sampling efficiently arising from slower $T_1$-contrast generation and partly static nature of underlying biophysical models.

ASL requires two inversion-recovery cycles for absolute CBF estimates. Due to the labeling duration and a PLD of about 1.7s before the readout module can be played out, label and control condition will each be in the order of 2s. Hence, effective temporal resolutions are rarely shorter than 4s. The VASO contrast relies not only on inflowing blood, but is further sensitive to blood magnetization that is located in the imaging slice during inversion and blood magnetization that flows into the imaging slice after inversion. Hence, it does not require a PLD, and inversion times can be in the range of 0.9–1.3s at 7T. Including a typical readout module, the temporal resolution is in the range of $T_R=1.5\mathrm{s}$. However, for interleaved VASO/BOLD acquisition schemes, as in the high-field optimized SS-SI-VASO method (Huber et al., 2014b), the effective temporal resolution doubles to about 3s. The calibrated fMRI sampling efficiency inherits the conditions for ASL or VASO and is additionally limited by a conserved interaction of vascular compartments. Since arteries, veins, CBF and CBV have heterogeneous response-time courses across cortical depth, ${\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$ estimations are typically considered to reach a steady-state after 5–10s. Adaptation of dynamic calibration models (Hyder, 2010; Simon and Buxton, 2015) for depth-dependent applications might, thus, be interesting for improving temporal efficiency.

5.5. Availability of the method

The wider spread application if of non-BOLD depth-dependent fMRI is mostly limited by the availability of cutting-edge MRI hardware (7T), sequence software, and know-how of appropriate analysis methods. ASL has been implemented across species and vendors. However, advanced sequence versions that are optimized for the specific on-site configuration (e.g., 7T) are still not widely available and usually require considerable sequence-development efforts. Although there are no commercial product versions of the VASO sequence (despite it’s relatively simple basic sequence design with inversion preparation), it has been independently implemented at 7T in about 6–10 research labs. The most-limiting additional factor for calibrated fMRI is the presence of suitable gas-breathing equipment in addition to the availability of ASL or VASO sequences. This limitation could be addressed, however, with alternative calibrated BOLD approaches that do not require breathing of special gases (Blockley et al., 2015a).

6. Recommended procedures of non-BOLD laminar fMRI

6.1. Recommended field strength

With increasing spatial resolution, the relative contribution of thermal noise increases whereas that of physiological noise (and signal) decreases. Experimentally, submillimeter voxels are in the thermal-noise-dominated regime and, therefore, any gain in image SNR is of crucial importance for laminar fMRI. Thus, we recommend field strengths of 7T or above for non-BOLD-fMRI. In addition, the longer $T_1$ allows use of longer $T_I$, resulting in amplified perfusion-related signal in ASL (Gardener et al., 2009) and VASO (Huber et al., 2014b; Jin and Kim, 2008b). While ASL and VASO applications in humans at 7T are now established, applications at 9.4T are just starting to get into the focus of current research efforts (Bause et al., 2016; Huber et al., 2017c).

6.2. Recommended contrast-generation methods

In a recent ASL consensus paper, pCASL is the recommended sequence of choice for most robust CBF estimation at 1.5 and 3T (Alsop et al., 2015). For 7T, however, there is no established consensus yet. At high field, RF transmit-field ($B_1^{+}$) homogeneity and SAR constraints currently limit the $T_R$ of pCASL to be approximately 5 s (Luh et al., 2013; Wang et al., 2015), reducing its applicability for fMRI. All published (in-plane) submillimeter human CBF studies have used FAIR (Duong et al., 2002; Ivanov et al., 2017; Pfeuffer et al., 2002) due to its lower SAR, relative robustness to field inhomogeneities and easiness of implementation; hence, currently, we also recommend it for depth-dependent fMRI in humans.

We further recommend VASO for laminar CBV-fMRI in humans. Because of longer blood $T_1$ at higher field, the SI-VASO (Jin and Kim, 2008b) or SS-SI-VASO modifications (Huber et al., 2014b) are advantageous over other VASO variants. Their tSNRs benefit from longer blood-nulling times, making them the most widely used variants for depth-dependent applications. We recommend VASO over other potential non-invasive CBV-sensitive MRI methods like VERVE (Stefanovic and Pike, 2005), hyperoxia-driven CBV methods (Blockley et al., 2011), magnetization transfer methods (Kim and Kim, 2005), and $T_2$-relaxation-dependent methods (Piechnik et al., 2009), because of its relatively high sensitivity that has been tested in humans at high resolutions.

Calibrated fMRI requires measurements of BOLD signal, CBF and CBV. Ideally, all three parameters should be measured simultaneously (Cheng et al., 2016, 2014; Krieger et al., 2014b; Yang et al., 2004). However, most human studies have relied on combined ASL and GE-BOLD acquisitions and Grubb’s relation (Eq. 3). Alternatively, ASL may be substituted by VASO (Eq. 4), and an assumption about ${\mathrm{C}\mathrm{B}\mathrm{V}}_0$ can be made based on known histological distributions. At 7T or above, where the ASL stability is increasingly limited by constraints in inversion coverage and efficiency (Driver et al., 2012; Krieger et al., 2014a, 2014c), the latter approach can be advantageous (Huber et al., 2015, 2014a) and was used in the only human laminar-dependent study published so far (Guidi et al., 2016b). The larger SNR in VASO compared to ASL can help to minimize non-linear error propagation in the Davis model. Until other evidence becomes available, we recommend the interleaved VASO/BOLD acquisition scheme (see above) for depth-dependent ${\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$ estimation in humans. Again, please note that, when using VASO in calibrated BOLD, different Grubb coefficients must be used for modeling the coupling between venous CBV and CBF as opposed to the coupling between the total CBV and CBF (Guidi et al., 2016b).

6.2.1. Recommended readout strategy

Since non-BOLD mechanisms are not based on susceptibility effects but rather manipulations of longitudinal magnetization, it is advantageous to choose $T_E$ as short as possible (Hetzer et al., 2011). In recent years, best readout practices for high-resolution non-BOLD-fMRI have been investigated for ASL (Ivanov et al., 2016a, 2016b) and VASO (Huber et al., 2016b, 2016c). Based on these cutting-edge studies, we recommend multi-shot 3D readouts for depth-dependent application of CBF- and CBV-fMRI. It must be noted that 3D-readouts (instead of 2D-readouts) are recommended particularly in the context non-BOLD contrasts at high fields with cortical-depth resolution. Their moderate SAR constraints (Poser et al., 2010) and higher relative image SNR (Huber et al., 2016c) make them favorable for the setup of non-BOLD contrasts.

Alternative promising readout strategies for non-BOLD fMRI could be 3D-TFE/TFL (Cheng et al., 2016; Hua et al., 2013) or 3D-GRASE (Poser and Norris, 2009). In particular, 3D-GRASE readouts offer great potential in BOLD-sensitive layer-dependent fMRI (e.g. (De Martino et al., 2013; Kemper et al., 2017, 2016, 2015; Zimmermann et al., 2011)). However, the potential of 3D-GRASE included in the various non-BOLD contrast fMRI approaches has not yet been shown for layer-dependent fMRI.

Since the application of parallel imaging acceleration and associated tSNR reductions are hardly accountable at already tSNR-’starved’ resolutions, we recommend to use small FOVs (Schluppeck et al., 2017) with minimal acceleration.

6.2.2. Recommended voxel size

The required spatial resolution of fMRI for appropriate laminar analysis is expected to be comparable in non-BOLD and conventional GE-BOLD-fMRI. We believe that for cortical-depth-dependent interpretations, a minimum of three different depth intervals is necessary. Hence, the recommended voxel size should be below 0.8 mm in areas where the thickness¹ of the cortical ribbon is in the range of 2 mm (e.g., S1 and V1), and it should be below 1 mm for thicker cortical brain areas (e.g., M1).

In some of the previous non-BOLD depth-dependent fMRI studies in humans, these requirements of more than three voxels across cortical depth could only be met with sufficient tSNR by means of using anisotropic voxels (0.7–0.8 mm in-plane and 1.5–1.8 mm slice thickness) and/or imaging particularly thick cortices (M1, 4 mm) (Guidi et al., 2016b; Huber et al., 2016c). Following the animal literature, such thick slices could be achieved without loss in resolution across cortical depth by placing the imaging plane perpendicular to the cortex. This, however, requires careful position planning and retrospective quality checks (see supplementary material of (Goense et al., 2012a)). When imaging cortical areas with curvature patterns along more than one dimension, bigger slice thickness cannot be applied and isotropic voxels are recommended.

Note that compared to pioneering animal work, depth-dependent fMRI data in humans are mostly acquired with coarser resolutions than 0.5 mm and they are challenged with stronger cortical folding patterns. This makes it more difficult to directly extract cortical laminar profiles of activation in humans compared to animals. Hence, the development of robust analysis methods is an intense area of current research (Kashyap et al., 2017; Kemper et al., 2017; Polimeni et al., 2017). The most used method included the layering to be applied in a finer grid compared to the fMRI data. Many methods involve averaging over an extended ROI, and some methods include spatial un-mixing to obtain laminar information in humans.

6.2.3. Recommended respiration task in calibrated BOLD

Since calibrated fMRI uses VASO and/or ASL, we suggest the same sequence-related parameters as discussed above. Additionally, we sought to provide recommendations about the experimental setup of the calibration experiment.

To estimate $M$, suitable manipulations include hypercapnia (Davis et al., 1998), hyperoxia (Chiarelli et al., 2007b), breath hold (Kastrup et al., 1999), or the Valsalva maneuver (Wu et al., 2015). Arguably best established and recommended is mild hypercapnia induced by breathing of CO₂-enriched air (e.g., 4–5% CO₂), which can be achieved with premixed gases and a simple non-rebreathing mask (Blockley et al., 2013). It provides a reasonably high CNR with sufficient robustness and is the most widely employed approach for depth-dependent fMRI with breathing manipulation (Bohraus et al., 2011; Guidi et al., 2016b; Yen et al., 2012). More complex respiratory challenges involving graded hypercapnia and hyperoxia often applied with more sophisticated, computer-controlled gas-delivery systems (Mark et al., 2011, 2010; Slessarev et al., 2007; Wise et al., 2007), however, may be advantageous as well.

7. Value of non-BOLD-fMRI: When it should be applied

Because of the high sensitivity of the GE-BOLD response, it has been the method of choice for experiments in humans, including high-resolution depth-dependent fMRI. Alternatively, SE-BOLD-fMRI can be utilized with higher spatial specificity due to lower sensitivity to ascending and pial veins albeit with potential reduced signal sensitivity. Nevertheless, both GE- and SE-fMRI are non-quantitative markers of neuronal activity. In specific applications, when the neurovascular coupling is suspected to be altered, additional non-BOLD contrasts are required for modeling and interpretation of fMRI data. These cases of altered neurovascular coupling include comparisons of regions with different baseline physiology (Guidi et al., 2016b), plasticity (Krieger et al., 2014a), aging (Ances et al., 2009; Gauthier et al., 2013; Liu et al., 2013; Mohtasib et al., 2012), neuropharmacology (Zaldivar et al., 2014), neural inhibition (Goense et al., 2012a; Huber et al., 2014a), or in the context of clinical applications. We specifically believe that clinical applications, with suspected deviations of the neuro-BOLD coupling, will be important to facilitating a more wide-spread application of non-BOLD layer-dependent fMRI. However, since the field of non-BOLD human layer-fMRI is still in an early phase, established clinical applications are still scarce. Potential future examples of layer-dependent clinical non-BOLD fMRI studies might include the following:

1. Modified cortical circuitry after peripheral nerve damage where plasticity changes affect the onset of vessel dilations from feed-forward driven layer II/III to cortico-cortical driven layers (Yu et al., 2014).

2. Development disorders with abnormal laminar cytoarchitecture and cortical disorganization of neurons, such as in autism (Stoner et al., 2014).

3. Investigation of schizophrenia. In recent studies bridging the macroscale and microscale brain architecture (van den Heuvel and Yeo, 2017), it was shown that changes of long-range connectivity in schizophrenia patients are associated with layer III pyramidal neuron architecture (van den Heuvel et al., 2016). Due to the reduced GM vascularization in schizophrenia patients (Hua et al., 2016), however, uncalibrated BOLD fMRI has a reduced interpretability and should be accompanied with non-BOLD fMRI.

Since baseline vascular physiology is different across cortical depths, coupling of CBF, CBV, and energy consumption can deviate and alter the BOLD response (Devor et al., 2008; Goense et al., 2012a). Hence, it has been suggested that for appropriate interpretation of depth-dependent BOLD results, they should be accompanied with CBF- and CBV-based fMRI (Goense et al., 2016). The, thereby, obtained additional information of neurovascular coupling is particularly important for interpretations of negative BOLD and neuropharmacological data.

8. Contrast-independent methods to improve layer specificity of fMRI

In the recent years, there have been few cortical depth-dependent application studies that infer layer-dependent activity information using GE-BOLD signal (e.g., (Dumoulin et al., 2017; Moerel et al., 2017; Uludaǧ and Blinder, 2017) and references therein). These studies address questions about feed-forward vs. feedback driven activity in human auditory and visual cortex by means of the BOLD signal (De Martino et al., 2015; Fracasso et al., 2016; Kok et al., 2016; Muckli et al., 2015; Olman et al., 2012; Polimeni et al., 2010). However, because of the aforementioned vascular biases of BOLD fMRI with regards to specificity and quantifiability, these studies do not interpret the fMRI signal directly in single-condition paradigms, but instead they try to control for vascular biases by the applications of differential conditions. As described in (Ugurbil, 2016), single-condition studies allow investigation of the task-induced activity vs. rest, while differential condition studies can only investigate the changes in activity between two tasks. However, since also the differential BOLD signal changes can be contaminated by baseline vascular biases as well (Kashyap et al., 2017; Markuerkiaga et al., 2016), depth-dependent fMRI responses have been increasingly investigated by means of computational models across differential conditions. Instead of direct fMRI signal change interpretations, the computational approaches are based on other data processing outcomes, such as, modulations in the population receptive field properties (Fracasso et al., 2016), modulations of population tuning curve widths (De Martino et al., 2015), or modulations in task classification accuracy (Muckli et al., 2015). The limitations of these approaches are that these measures are tightly reliant on the underlying biophysical model assumptions. These model assumptions, however, may be influenced by: the variations in the noise models across cortical depths (Fig. 1B, 2D), variations in the hemodynamic response (Fig. 2C), and variations in the underlying physiology across cortical depths (Fig. 1B–C), all of which together hamper direct interpretation.

Since non-BOLD methods discussed here can show cortical-depth dependent activity with reduced vascular biases, they might permit the use of both single-condition and differential condition paradigms for the study of layer-dependent functional profiles, which in turn enlarges the set of experimental setups available.

9. Conclusion

In the field of neuroimaging, non-BOLD-fMRI is just starting to be customized for cortical-depth-dependent applications. Techniques have been established in the animal literature but are not yet as widely applied in humans. Few research groups, however, are pioneering high-resolution ASL, VASO, and calibrated fMRI in humans for depth-dependent CBF, CBV, and ${\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$ quantifications, respectively. These methods already provide promising initial results. In ‘SNR-starved’ submillimeter fMRI, they are still challenged by smaller sensitivity compared to the GE-BOLD response. However, they can provide improved specificity, quantifiability and, hence, more straightforward physiological interpretability of cortical profiles. We believe that future application of non-BOLD-fMRI will not only be a useful alternative for depth-dependent investigations, but will provide valuable insights into neurovascular coupling and cognitive processes in the brain.

Table 2.

Summary of recommended scan parameters.

Recommended scan parameters	ASL (CBF)	VASO (CBV)	Calibrated fMRI (CMRO2)
Field strength	7T	7T	7T
Sequence	FAIR with 3D readout	SS-SI-VASO with 3D readout	VASO/BOLD with 5% CO2
Resolution	Voxels should be sub-dividable from at least three different depth intervals: 0.75 mm for thin cortices such as V1, 1 mm for thick cortices such as M1
$T_R$	2 s (label) and 2s (control)	1.5 s (blood nulling) and additional 1.5 s for interleaved BOLD acquisition in SS-SI VASO	approx. 5–10 s (after functional steady-state is reached)
$T_E$	as short as possible		optimal for VASO and BOLD
Acceleration	as minimal as possible in thermal-noise dominated non-BOLD fMRI

Highlights.

• Non-BOLD-fMRI methods are reviewed for layer-dependent application in humans.

• ASL, VASO, and calibrated fMRI are promising candidates.

• Representative depth-dependent ASL/VASO/calibrated fMRI results are shown.

• ASL/VASO/calibrated fMRI are less sensitive but more specific than GE-BOLD.

Abbreviations:

2DG

2-deoxy-d-glucose

3D

three-dimensional

AAT

arterial transit time

ASL

arterial spin labeling

BOLD

blood oxygenation level-dependent

dHb

deoxyhemoglobin

CA

contrast agent

CASL

continuous ASL

CNR

contrast-to-noise ratio

CSF

cerebrospinal fluid

EPI

echo planar imaging

FAIR

flow alternating inversion recovery

FLASH

fast low-angle shot

fMRI

functional magnetic resonance imaging

FOV

field of view

GE

gradient echo

GM

grey matter

GRASE

gradient and spin echo

IRON

increased relaxation for optimized neuroimaging

LFP

local field potential

M1

primary motor cortex

MION

monocrystalline iron oxide nanocompounds

MR

magnetic resonance

MRI

magnetic resonance imaging

MT

magnetization transfer

PASL

pulsed ASL

pCASL

pseudo-continuous ASL

PET

positron emission tomography

PICORE

proximal inversion with a control for off-resonance effects

PLD

post-label delay

QUIPSS II

quantitative imaging of perfusion using a single subtraction, second version

RBC

red blood cell

RF

radiofrequency

rs

resting state

S1

primary somatosensory cortex

SAR

specific absorption rate

SE

spin echo

SI-VASO

slab-inversion VASO

SMS

simultaneous multislice

SNR

signal-to-noise ratio

SS-SI-VASO

slab-selective-single-inversion VASO

TFE

turbo field echo

TFL

turbo FLASH

USPIO

ultra small superparamagnetic iron oxide

V1

primary visual cortex

VASO

vascular space occupancy

tSNR

temporal SNR

Mathematical symbols

$B_1^{+}$

RF transmit magnetic field amplitude

CBF

cerebral blood flow

CBV

cerebral blood volume

${\mathrm{C}\mathrm{M}\mathrm{R}}_{\text{Glc}}$

cerebral metabolic rate of glucose

${\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$

cerebral metabolic rate of oxygen

$f$

flow in units of 1/s

$M$

calibration constant

$n$

coupling constant for the relation between CBF and ${\mathrm{C}\mathrm{M}\mathrm{R}}_{\mathrm{O}2}$

OEF

oxygen extraction fraction

$R_2^{\mathrm{*}}$

effective transverse relaxation rate

$R_2^{\prime }$

magnetic-field-inhomogeneity-induced transverse relaxation rate

$S$

signal

$\mathrm{\Delta }{S}_{\text{ASL}}$

ASL signal difference between control and labeling condition

$\mathrm{\Delta }{S}_{\text{BOLD}}$

BOLD signal change during to stimulation task

$T_1$

spin-lattice relaxation time

$T_2^{\mathrm{*}}$

effective spin-spin relaxation time

$T_E$

echo time

$T_I$

inversion time

$T_R$

repetition time

$\alpha$

Grubb’s constant

$\beta$

exponent in the power-law dependence of $R_2^{\prime }$ on $\left[\mathrm{d}\mathrm{H}\mathrm{b}\right]$

$\lambda$

blood-brain partition coefficient

0

index indicating a baseline value

$B$

index indicating blood

$\mathit{ex}$

index indicating the extravascular compartment

$\mathit{in}$

index indicating the intravascular compartment

$t$

index indicating the total blood compartment

$v$

index indicating the venous compartment

$\mathrm{\Delta }x$

change of the variable $x$

[X]

concentration of compound X