Functional neuroimaging studies seem to have a life cycle where a small scale study calls for its findings to be repeated with more participants. This larger scale study then calls for it to be repeated at higher resolution. Facings cost restraints the researchers using higher resolution carry out the study on a small number of participants and then call for it to be repeated on a larger scale.
Here are a few thoughts about resolution in functional neuroimaging.
Are we reaching the limits of what BOLD FMRI can reveal about brain function?
The ability to harness the phenomenon of magnetic resonance has led to a revolution in imaging in research and medicine. Magnetic Resonance Imaging (MRI) is dependent upon the detection of energy emitted as electromagnetic waves from excited nuclei as they relax after a period of excitation.
Oxygen in the brain is transported by blood which contains hemoglobin (Hb). Deoxyhemoglobin (dHb) acts as an “endogenous paramagnetic contrast agent” (Pauling & Coryell, 1936). Changes in dHb lead to changes in the MR signal intensity which underpins the use of the blood oxygen level dependent (BOLD) contrast. Distortion of the local magnetic field by dHb effects the relaxation time of hydrogen nuclei, reducing the T2 and T2* signals (Kim & Ogawa, 2012). Likewise an increase in bloodflow increases the proportion of oxyhemoglobin and generates the BOLD signal. This measurable effect can be recorded and referred to as functional magnetic resonance imaging (fMRI).
Increased neuronal activity can be associated with an increase in cerebral blood flow (CBF) – an increase in blood flow which transports hemoglobin – as metabolic demands require the delivery of oxyhemoglobin for the oxidation of glucose. A change in the BOLD signal is facilitated by an increased CBF which arises due to increased neuronal activity (Raichle, 1987).
As such the BOLD response is a measure of metabolic activity and not a direct measure of neuronal activity. fMRI offers a non-invasive methodology to study the brain which has advantages as well as limitations. An fMRI study must have an experimental method and an MRI scanner which both have bounds to their capabilities. In addition the physiological response of the brain will be limited by the response of the vascular system – the hemodynamic response function (HRF).
Where the resolution of the imaging system is not the limiting factor, physiological factors must be accounted for to determine the point that the structural architecture of the brain limits inferences that can be made of the organ’s functional response.
Being an indirect measure of neuronal activity, the temporal resolution fMRI will be limited by the HRF. In this respect BOLD based techniques are limited compared to magnetoencephalography (MEG) or electroencephalography (EEG) which can measure changes in current during synaptic transmission.
Physiological Factors
Changes in the venous concentration of dHb underpins the use of BOLD contrast based image acquisition. The BOLD signal can be considered a measure of dHb in a voxel. Concentrations of dHb vary around the brain due to the functional architecture of the vascular system. The effect of a high relative concentration of dHb in a voxel mapping a large vein is to limit spatial resolution (Ogawa et al., 1993).
Both intravascular (IV) and extravascular (EV) components contribute to the BOLD signal. During the time course of MR excitation, water is exchanged intravascularly between deoxygenated blood and plasma. Water molecules may diffuse with different field orientations between locations. If this movement in space occurs faster than can be temporally resolved (i.e. the echo time taken being the integration time at that discrete spatial location) aliasing may occur resulting in a reduction of the T2 signal recorded from the IV component. In addition if many blood vessels are present within a voxel, a phase dispersion will be caused by their random orientations resulting in a decrease of T2* (Kim and Bandettini, 2010).
Molecules of water in extravascular blood may diffuse by 10 – 20µm during a sampling time of 50ms, likewise producing temporal aliasing (Kim & Bandettini, 2010). As this measure is integrated within the MR signal this effectively places a spatial constraint upon the sample – no distinction between discrete elements could be made below the extent of the diffusion by water molecules. The magnitude of this effect is dependent upon the size and orientation of the blood vessel that EV water is proximal to. As such, field gradients produced by dHb are less steep near a larger blood vessel than one of smaller diameter. This leads to a local averaging of dephasing effects due to EV components near larger vessels compared to a dynamic averaging from EV water close to a small vessel (See Figure 1.) (Kim & Ogawa, 2012).
Dephasing of spins in water in IV and EV locations as they diffuse is the “primary mechanism for the BOLD signal” (Huettel et al., 2014). If this mechanism is affected by alignment of the blood vessel, containing dHb, to the static magnetic field (B0) and the extent over which field distortions are exhibited due to size of the blood vessel this may place a fundamental limit on the spatial resolution obtainable by BOLD fMRI.
A spatial constraint is placed upon the sampled signal received as no distinction between discrete elements could be made below the extent of the diffusion by water molecules.
As diffusion of water molecules is affected by dHb, the ultimate spatial resolution of BOLD based imaging depends at what level cerebral blood flow is regulated (Kim & Ogawa, 2002). If CBF is controlled at the level of the capillary a higher spatial resolution would be theoretically possible than if it were found to be controlled at the arteriole level.
A voxel size at the arteriole level (10µm) would require a large magnetic field that if technically possible would need to be determined to be safe for humans. Even so, when considering a voxel of 3mm in dimension this gives an region for improvement of two orders or magnitude, with 27 million 10µm voxels, (3*10-3/10*10-6)3, occupying the same volume.
Overcoming Physical limitations
With this difference it would seem that BOLD fMRI appears to be currently limited in spatial resolution by the physical capabilities of hardware than the physiological response. Several avenues may be considered for the improvement of the current physical capabilities of BOLD fMRI; increasing the strength of the static magnetic field, B0, improvements in methods of excitation and reception, both in terms of hardware and signal processing to enable either faster or more precise measurements without obtaining a signal modulated by noise.
Field Strength
A seemingly straightforward way to increase the spatial resolution of fMRI is to increase the field strength of electromagnetic field in which images are obtained. Increasing the field strength would facilitate the isolation of more spin frequencies due to the Lamor equation where ω is the Lamour frequency given by γ, the gyromagnetic constant.
ω = γ . B0
The field of view required to image a (whole) brain would be the same, regardless of the field strength employed. It is dependent on the dimensions of the subject being imaged. However the number of elements giving the matrix size would increase, increasing the number of voxels per imaged volume. While ostensibly providing an increase in spatial precision, as noted above, the reception of the sampled property is a passive one. The amplitude of signal emitted by the subject will still be distributed over the same volume, but now that volume has been subdivided into a greater number of voxels. The noise of the system interfering with the actinic signal however becomes much larger in proportion. The decrease in signal to noise ratio (SNR) therefore reduces the image contrast.
While technological obstacles may be overcome in the future which currently limit the strength and homogeneity of the static magnetic field, leading to increased spatial resolution, there may be a limit to the strength of the field that humans can tolerate. Atkinson et al. (2010) conducted MRI at 9.4T. Although they reported that “vital signs and cognitive function” were not affected they did report that participants experienced a range of sensations; “temperature change, a metallic taste, vertigo, lightheadedness (sic), nausea, muscle twitching or tingling, visual perception of one or more flashes of light, anxiety and sleepiness.” The limit at which a high strength magnetic field is safe for humans to be exposed to may be different from one that is tolerable. The limit on spatial resolution due to B0 may eventually be determined by the limits that a participant can stand within the scanner. If effects due to B0 are noticeable it should be questioned at level it impinges upon the participants ability to accurately respond during an experiment – despite the verbal assurances reported – and so undermine the data being obtained.
Signal processing and hardware improvements
In terms of imaging metrics at higher magnetic fields the IV contribution from large vessel decreases while the EV component increases (Ogawa et al., 1998). The EV component from large vessels remains significant regardless of magnetic field strength. The dephasing effect – most prominent when large vessels are orthogonal to B0 – can be refocused by the use of a 180o RF pulse. This spin-echo technique would reduce the aliasing affects of the EV component compared to gradient-echo techniques at higher field strengths. Spin-echo techniques are inherently less sensitive than gradient-echo techniques so contrast would be sacrificed (Lee et al., 1999). Additionally they require a longer sampling aperture which increases the time taken to acquire a brain volume.
Employing a spin-echo sequence with a short echo time has been found to produce results that deviate from the expected BOLD signal. This implies additional interactions in the EV component that could be harnessed as an imaging pathway. “Signal enhancement by extravascular water protons” or SEEP may provide an imaging methodology near large blood vessels or cavities where BOLD imaging provides weaker results (Figley et al., 2010).
Improvements in technology may facilitate improvements to spatial resolution such as head coils that contain shimming apparatus in addition to radio frequencies transmitters and receivers. The increased proximity to the brain of the shimming circuit locally reduces inhomogenities in the magnetic field. In addition to increased sampling precision, greater clarity in reception may lead to an increase SNR (Huettel et al., 2014).
Integration of technology into head coils has allowed the use of multi-channel imaging to acquire sampled MR data in parallel. Using an array of coils, of characterized sensitivity, the fall off in reception can be effectively triangulated by solving a matrix equation of the received frequency domain signal. The signal can be interpreted with reduced Fourier encoding and a reduction of associated imaging artifacts. Parallel imaging methodologies such as SENSE and SMASH may allow a greater matrix size and number of slices, therefore reducing voxel size and increasing spatial resolution over EPI slice acquisition. However a longer sampling aperture required may lead to temporal aliasing. By using a spiral sampling pattern over an array of coils a reduced echo time can be employed. The trade off for achieving this reduction is greater complexity in decoding the received signal. However, as the information has been read-out from the head coil, this information can be processed downstream. As such it does not impinge upon the progress of the scanning session, with the equipment free to take the next measurement in the series (Zahneisen et al., 2014).
Image quality may be improved by better signal processing of the raw data. Improving SNR facilitates the improvements in temporal as well as spatial resolution. If the BOLD technique can be applied with improvements to these parameters advances can be made without upgrades to hardware (Lee et al., 2016) have proposed the use of Annihilating Filter-Based Low Rank Hankel (ALOHA) Matrixes to correct for ghosting artifacts within echo-planar imaging sequences. By employing a matrix to interpolate the signal in frequency space the disparity between odd and even EPI data is resolved eliminating aliasing in the spatial domain. The elimination of such artifacts caused by B0 inhomogeneities and eddy currents by the pulse sequence may lead to faster parallel scanning both in echo and repetition time.
Enhanced Experimental Methodologies
Advances in interrogating data obtained by fMRI may lead to progress within the field. By using innovative techniques to recover information obscured within the signal, constraints of current physical factors may be pushed back.
Multi-Voxel Pattern Analysis (MVPA) looks for correlations and patterns between voxels. Traditionally fMRI techniques have looked for evidence of areas of activation. With this broad brush approach the loss of spatial precision by smoothing the acquired data by comparison with readings from neighbouring voxel was not critical. The smoothing process negates the effects of noise modulating the received signal.
The MVPA approach requires experimental data, generated by response to a stimulus, is divided into two classes; a training set and a testing set(s). Data within the training set is examined for correlation between values. The robustness of these correlations are then tested by comparison with the testing set(s). A strong and robust correlation should enable a researcher to classify the class stimuli from a modelled data set the cortical activation evoked. The designation of training and testing sets can be systematically rotated to refine the accuracy of this model.
Negating the effects of noise by covariation leads to the preservation of spatial precision due to the avoidance of smoothing. However the recent research by Dubois et al. (2015) may indicate that this retention of detail by MVPA is too course to capture some neural responses. Comparing activation in an area corresponding to the temporal lobe against findings from single-unit recordings in macaques they found correlation between methods when presenting stimuli of (human) individual’s faces presented at different viewpoints. However single-unit recording revealed cortical activity discerning differences between individuals’ identity which was not detected by the fMRI technique (See Figure 2).
The authors attributed the failure of MVPA to resolve activity in the second case to the “sparseness” of the neural population. That MVPA could not resolve the activity due to the widely distributed nature of neurons in the network – not necessarily its population. Contrasted against areas of discrete activation revealed by functional localisation the HRF generated is summed over too large a volume to be resolvable. It is interesting to consider the possibility that the presence of functional networks are undiscovered due to their distributed nature (both through failing to elicit a resolvable functional response or distinguishable structural architecture by other techniques such as diffusion tensor imaging) and how a “discovery” of such sparse pathways would impact our understanding of the brain’s operation.
Figure 2 – reproduced from Dubois et al. (2015)
Recent attempts by Norman-Haignere et al. (2015) to overcome the coarseness created by the physical limitations of the scanning procedure have focused on the specification of canonical responses of a neural population. The rationale behind this method is that a voxel contains hundreds of thousands of neurons. By isolating neuronal subpopulations within a voxel to infer “response dimensions from structure in the data, rather than testing particular features hypothesized to drive neural responses”.
The production of inferred canonical responses to 165 natural sounds revealed six components; four specifying the acoustic nature of the sound and two whose response were not tied to the physical production of the stimulus. The two latter components exhibited selectively depending on the nature of the sound; whether the stimulus was music or speech. This finding has revealed the presence of distinct cortical pathways for music and speech based on the BOLD. In addition by applying this “hypothesis-free” voxel decomposition post facto researchers may tell whether the stimulus was of a musical or spoken nature dependent on the neural response without knowledge of the stimulus as long as it was tuned to a specified canonical response (See Figure 3).
Figure 3 reproduced from Norman-Haignere et al. (2015)
Conclusion
While fMRI has opened up many exciting avenues for scientific discovery it is not a panacea. Any technique will fail when it is applied to a task that it was not designed to do.
fMRI is not a direct measure of neural activity and so its temporal resolution will be ultimately limited by the HRF which may make it unsuitable for some tasks when compared with techniques such as MEG or EEG. Parallel imaging and downstream processing decoding of complex signal may improve the temporal resolution of fMRI towards this bound, but with the benefit of providing better (whole brain) spatial resolution over alternative techniques, so opening up new avenues for exploring brain function.
The physiology of the human brain suggests that there is significant extent for improvement of spatial resolution from that offered by current BOLD imaging standards. However the field strengths needed to facilitate this resolution may not be tolerated by the human body. In addition would any greater resolution actually give us any greater insight into the function of the brain? Neuroanatomist Valentino Braitenberg notes, “it makes no sense to read a newspaper under a microscope”. *
Just like choosing an appropriately scaled map the appropriate FOV and resolution depends on the experimental methodology and the underlying structural architecture being studied. An experimenter may choose to obtain fine data of a particular landmark during a study and repeat to obtain a wider view of the subject (or vice versa). The ability to work at high resolutions may yield new insight on the workings of small cortical structures but may hinder research on broader areas due to the additional load of process the excess data acquired.
Improvements to current BOLD based imaging standards may come on the form of improved signal processing and experimental methodologies. Parallel imaging technique may facilitate faster TR times while maintaining a strong SNR. With increased complexity comes increased processing, which can be handled downstream of the scanning session, provided the load of information transferred from the head coil does not become a limiting factor to scanning.
Signal processing techniques such as ALOHA and data analysis methodologies such as MVPA or the profiling of canonical responses, where applicable, may lead to more information being extracted from a data set.
While BOLD fMRI may ultimately be too coarse a technique to reveal some neuronal populations revealed by techniques such as single cell recordings, it provides a safe, non-invasive repeatable technique. There are limits to the capabilities that imaging of the BOLD response can reveal about brain function but with advances in the areas described it does not seem likely that the limits which the technique can reveal about brain function have been reached.
Photo credit: © Amlani
* As seen in Logothetis, N. K. (2008). What we can do and what we cannot do with fMRI. Nature, 453(7197), 869-878.
References
Atkinson, I. C., Sonstegaard, R., Pliskin, N. H., & Thulborn, K. R. (2010). Vital signs and cognitive function are not affected by 23-sodium and 17-oxygen magnetic resonance imaging of the human brain at 9.4 T. Journal of Magnetic Resonance Imaging : JMRI, 32(1), 82-87
Dubois, J., de Berker, A. O., & Tsao, D. Y. (2015). Single-unit recordings in the macaque face patch system reveal limitations of fMRI MVPA. The Journal of Neuroscience : The Official Journal of the Society for Neuroscience, 35(6),
Figley, C. R., Leitch, J. K., & Stroman, P. W. (2010). In contrast to BOLD: Signal enhancement by extravascular water protons as an alternative mechanism of endogenous fMRI signal change. Magnetic Resonance Imaging, 28(8), 1234-1243.
Huettel, S. A., Song, A. W., & McCarthy, G. (2014).Functional magnetic resonance imaging (Third Edition). Sunderland, Massachusetts, USA: Sinauer Associates, Inc. Publishers. a: PP. 238-243, b: PP. 470-471.
Kim and Bandettini (2010). Seen in BOLD fMRI: A Guide to Functional Imaging for Neuroscientists. Editors Faro, S. H., & Mohamed, F. B. New York, USA: Springer Science & Business Media. PP 3 – 21.
Kim, S., & Ogawa, S. (2002). Insights into new techniques for high resolution functional MRI. England: Elsevier Ltd. PP. 607-615.
Kim, S., & Ogawa, S. (2012). Biophysical and physiological origins of blood oxygenation level-dependent fMRI signals. Journal of Cerebral Blood Flow and Metabolism : Official Journal of the International Society of Cerebral Blood Flow and Metabolism, 32(7), 1188-1206.
Kim, S., & Ugurbil, K. (2003). High-resolution functional magnetic resonance imaging of the animal brain. Methods,30(1), 28-41.
Lee, J., Jin, K.H., & Ye, J.C., (2016). Reference-Free Single-Pass EPI Nyquist Ghost Correction Using Annihilating Filter-Based Low Rank Hankel Matrix (ALOHA). Magnetic Resonance in Medicine, EPUB ahead of publication: http://onlinelibrary.wiley.com/doi/10.1002/mrm.26077/epdf, Retrieved: 20/02/2016.
Lee, S., Silva, A. C., Ugurbil, K., & Kim, S. (1999). Diffusion‐weighted spin‐echo fMRI at 9.4 T: Microvascular/tissue contribution to BOLD signal changes. Magnetic Resonance in Medicine, 42(5), 919-928.
Norman-Haignere, S., Kanwisher, N.G., McDermott, J.H. (2015) Distinct cortical pathways for music and speech revealed by hypothesis-free voxel decomposition. Neuron, 88, 1281-1296.
Ogawa, S., Lee, T. M., Nayak, A. S., & Glynn, P. (1990). Oxygenation-sensitive contrast in magnetic resonance image of rodent brain at high magnetic fields. Magnetic Resonance in Medicine, 14(1), 68.
Ogawa, S., Menon, R. S., Tank, D. W., Kim, S. G., Merkle, H., Ellermann, J. M., & Ugurbil, K. (1993). Functional brain mapping by blood oxygenation level-dependent contrast magnetic resonance imaging. A comparison of signal characteristics with a biophysical model. Biophysical Journal, 64(3), 803-812.
Ogawa, S., Menon, R. S., Kim, S., & Ugurbil, K. (1998). on the characteristics of functional magnetic resonance imaging of the brain. Annual Review of Biophysics and Biomolecular Structure, 27(1), 447-474.
Pauling, L., & Coryell, C. D. (1936). The Magnetic Properties and Structure of Hemoglobin, Oxyhemoglobin and Carbonmonoxyhemoglobin. Proceedings of the National Academy of Sciences of the United States of America, 22(4), 210–216.
Raichle ME (1987) Circulatory and metabolic correlates of brain function in normal humans. In: Handbook of physiology-the nervous system. Bethesda: American Physiological Society Vol V, pp 643–74.
Zahneisen, B., Poser, B. A., Ernst, T., & Stenger, A. V. (2014). Simultaneous multi-slice fMRI using spiral trajectories. Neuroimage, 92, 8-18.