Functional MRI can be accomplished with a high resolution gradient echo scan. The disadvantages of this approach are that the acquisition time can be many minutes. A further disadvantage is that as the resolution increases, the signal-to-noise (SNR) decreases. However, the data can be complex filtered back down to an equivalent lower resolution EPI like image to regain SNR. A simple subtraction can be performed rather than a correlation analysis. We also examine the role of using the complex data in the subtraction process rather than just the magnitude data. This approach may prove useful when studying activated tissue near tumors for example where the high resolution information may prove most useful. We test this approach using a conventional motor cortex fMRI paradigm involving finger tapping with and without the use of caffeine as an enhancer of the BOLD effect.
The conventional functional MRI (fMRI) map offers information indirectly about localized changes in neural activity because it reflects changes in blood oxygenation, not the actual neural activity. To provide neural basis of fMRI researchers have combined electrophysiology and various optical methods to show correlations between fMRI and surrogate signals associated with neural activity. But quantitative interpretation of "How much has the neural activity changed by?" still cannot be made from conventional fMRI data. The fMRI signal (S) has two partitions, one that describes the correlation between oxidative metabolism (CMRO2) and blood flow (CBF) which supports the bioelectric work to sustain neuronal excitability and the other is the requisite dilation of blood vessels (CBV) which is the mechanical response involved in removal of waste while providing nutrients. Since changes in energy metabolism is related to bioelectric work, we tested if spiking frequency of a large neuronal ensemble (v) in the cerebral cortex is reflected by local energy metabolism (CMRO2) in rat brain. We used extracellular recordings to measure dv/v and calibrated fMRI (using S, CBF, and CBV maps) to measure CMRO2/CMRO2 during sensory stimulation. We found that dCMRO2/CMRO2 ~ dv/v, which suggests efficient energy use during brain work. Thus calibrated fMRI can be used to provide data on where and by how much the neural activity has changed. We have probed the oxygenated environment of neural cells using fluorescence quenching methods. The localized oxygen partial pressure (pO2) measurements combined with quantitative measurements of oxidative metabolism (CMRO2) and blood perfusion (CBF) provide insights about oxygen back flux from brain to blood. The degree of oxygen back flux has bearings on the 'balloon' model, which is often used to describe the hemodynamic components of the stimulation-induced fMRI response. Since the results suggest that there is negligible oxygen back flux (from brain to blood), the oxygen transport process (from blood to brain) is believed to be far more efficient than assumed by the 'balloon' model. Recently we have also probed local temperature (T) changes in the brain and have begun to understand these changes with respect to quantitative changes in oxidative metabolism (CMRO2) and blood perfusion (CBF) during functional activation. The results suggest that the stimulation-induced temperature dynamics are heavily dependent on biophysical properties of heat transfer across different media and depend heavily on cooling and warming affects caused by blood flow and tissue metabolism, respectively. These combined multi-modal studies reveal the neuroenergetic basis of fMRI which is often an ignored aspect of the physiological makeup of the image contrast.
Several statistical approaches exist to compensate for the temporal smoothing effect inherent when using the BOLD response as a proxy for neural activation. Commonly used BOLD correction methods, such as convolving a stimulus function with a hemodynamic response kernel, inevitably make assumptions restricting the possible shapes of the BOLD response. Furthermore, the BOLD response shape is typically restricted so that only the response magnitude can vary spatially.
These assumptions were examined by fitting a range of parametric "shape" functions to voxel averaged BOLD response cycles using least squares estimation. The results imply that the shape of the BOLD response can vary spatially in a coherent fashion which, if ignored, could have implications on the detection and interpretation of activation patterns.
Significant progress has been made in last 10 years in terms of refining fMRI statistical analyses, acquiring empirical data on the relationship between BOLD signal and underlying physiology, and modeling these processes. Two questions addresses by this workshop include 1) How can we improve fMRI detection power, and 2) what do these signal changes mean in terms of hemodynamic, metabolic, and neuronal activity? In some (but not all) cases, a better understanding of the latter issue can inform statistical methods that define brain activation. One way to subdivide issues surrounding detection power and physiological modeling is to separately consider steady state and dynamic changes in fMRI signal.
fMRI between steady states: From a statistical viewpoint, long block designs have a minimal dependence on the hemodynamic response function. Pharmacological stimuli present an extreme form of the one-stimulus block design, where hemodynamic modeling is essentially irrelevant due to the slow evolution of neuronal activity. These cases limit certain options for analyses and place a premium on intrinsic sensitivity. From a modeling viewpoint, block designs and drug stimuli reduce sensitivity to BOLD transients, and help define the limits of interpretation in a simplified regime.
Numerous investigators, using the "hypercapnia calibration" methodology [1, 2], now have reported the relationship between steady state changes in CBF/CBV and CMRO2 [1-5]. The consensus appears to be that changes in CBF exceed those in CMRO2 by a factor between 2 and 3. These empirical data are not inconsistent with a diffusion-limited model of oxygen delivery  that includes capillary swelling, which was not included in the original model. I will argue that further refinements of both models and empirical data using fMRI techniques are limited by our uncertainties in physiological inputs.
The resting state BOLD relaxation rate amplifies changes in reactivity according to the local blood volume fraction and the magnetic field strength. The magnetic field dependence of the relaxation rate (and, hence, intrinsic BOLD sensitivity) can be investigated by 1) a "hypercapnia calibration" procedure, 2) comparisons between BOLD signal and fMRI based upon exogenous contrast agent, and 3) inferences based upon stimulus-induced changes in relaxation rates. These methods provide predictions for BOLD amplitude versus field strength. Empirical data show a regional coupling of BOLD and CBV signal changes, with a strong dependence of BOLD signal on resting state CBV [7, 8]. The BOLD dependence on resting state CBV represents a major impediment in terms of reliably and routinely translating BOLD signal to quantitative indices of neuronal activity.
Dynamic fMRI: Modeling dynamic fMRI data, such as event-related studies, requires a detailed understanding of transient features of the fMRI response and non-linearities that arise between the stimulus design and the measured output. It is now clear that a temporal mismatch between flow and volume is one of the major sources of BOLD transients. In both the anesthetized rodent [2, 9] and the awake non-human primate , the slow response of CBV is consistent with the time constant required to explain the BOLD post-stimulus undershoot. In each of these animal models, a detailed look at the temporal response of CBV shows 2 distinct time constants (much as BOLD signal appears to have one time constant for the dominant positive response, plus another slower time constant to describe the post-stimulus undershoot). In this section of the presentation, I will 1) review our empirical data on the responses of blood plasma and total hemoglobin , 2) describe models of this response  and discuss open questions about the physiological source of the flow-volume temporal mismatch, and 3) discuss the linearity of the CBV response, and implications for rapid event-related stimulus designs using BOLD and CBV contrast . For short or rapidly presented stimuli, attempts to derive CMRO2 or ascribe significance to fine temporal features of the time course are complicated by transit time effects.
In summary, statistical refinements may yield modest improvements in BOLD sensitivity for some paradigms. Some outstanding issues remain in terms of modeling. In general, certain experimental limitations, such as the difficulty in determining the BOLD baseline, will continue to hamper quantitative interpretations of BOLD signal changes in the routine experimental setting.
The talk will describe recent work developing a biophysical model linking the neural responses to stimulation, through the hemodynamic changes in blood oxygenation flow and volume, to the BOLD fMRI signal. In particular the major focus will be on the use of optical imaging spectroscopy and LDF measurements made concurrently with the BOLD and cbv-MRI measurements to examine the predictions of what is sometimes known as the Massachusetts General Hospital model of the BOLD signal (eg Boxerman, Davis, Hoge etc). Some even more recent work will be described in which IVIM crushing is used to explore the predictions of the 'Yablonskiy Haacke' (1994) model of the contribution of the extravascular static regime to the BOLD signal. The topic for discussion is this : despite the fact that the spectroscopy and the LDF (using Grubb's 'law') data is commensurate with the MRI measurements of the changes in blood volume, the measured BOLD signal is much larger than that predicted by the models (using generally accepted assumptions of baseline values) and the concurrent optical imaging measurements of changes in Hbr and Hbt.
Our aim is to conceptualize the complex physiology/pathophysiology that is involved in changes of oxygen in tissue and then apply advanced computational methods to develop a comprehensive physiological model that describes the distribution and changes of oxygen in tissue and the metabolic and signaling events associated with oxygen. This will be done using data from several different and complimentary methods for making measurements in vivo. Because the distribution of oxygen in tissues is very heterogeneous, even at cellular dimensions, such measurements and the resulting model are important but challenging tasks.
The need and opportunities for developing a comprehensive model for oxygen in tissues that is consistent with and validated by direct measurements, arose from studies that began as validation of EPR oximetry. As the "new" method, it was desirable to show that the measurements obtained with EPR oximetry "gave the same results" as other methods for measuring oxygen in tissues. We therefore initiated studies to make careful simultaneous or sequential measurements with EPR and one or more other modalities to determine the relationships between the results obtained with the various methods, taking into account the parameters on which the measurements are based.
As we began to carry out these experiments, however, we became acutely aware that the idea that we could do this via simple direct comparisons was an illusion. For example, it seemed logical to make direct measurements of oxygen in tumors simultaneously with EPR oximetry and the "gold standard", the Eppendorf Histograph. But the measurements are not really directly comparable even though they both measure the partial pressure of oxygen. This is because the volume measured with the EPR oximetry technique that we used is much larger than the volume probed with a single point with the Eppendorf. Even if we aggregate the volumes probed with the Eppendorf to make the total volume comparable to that with EPR oximetry, the Eppendorf measurements are spatially different and inevitably should record extremes of values that would not be recorded with the EPR method, because of the heterogeneity of the oxygen in real tissues, especially tumors. Therefore, even if both methods were technically perfect and valid, the results would be different. This conclusion, of course, applies to comparisons of essentially all types of measurements of oxygen and related parameters.
The relationship of measurements made with the BOLD effect to those made with other modalities is especially interesting, because the BOLD technique is widely available and it can be used to make measurements in virtually any part of an animal or human subject. The data obtained with BOLD, however, are very non-specific, reflecting principally the amount of deoxyhemoglobin. These data can be made more useful if they are combined with another type of related measurement, e.g. direct measurements of oxygen.
While data comparing results with two or more different modalities are valuable, these experiments have made us aware that it would be possible to develop a much more thorough and useful understanding of oxygen in tissue if we developed methods and models that can incorporate the different measurements into a physiologically based model that is based on the nature of the data from the different types of measurements. The oxygen concentration at any point is affected by the delivery, distribution, and consumption of oxygen locally, regionally, and systemically. These parameters are affected by many different processes including perfusion, diffusion, metabolism, the anatomy of the microcirculation, and the function of the macrocirculatory system. There are methods available to measure parameters that can be affected by most or all of the processes, but because each measured parameter is affected by multiple processes, multiple types of measurements are desirable. It also is desirable to have a logical basis to relate the measurements to each other; i.e., appropriate models of the processes. This would not only enhance the value of the data from the various types of measurements but, most importantly, also could lead to an optimized model that much more fully describes oxygenation in tissues, at levels ranging from the subcellular to the whole organism.
This is a challenging task, requiring input from several different disciplines, with multiple iterations to feed back the results of measurements into the model, then to modify the model appropriately and carry out measurements under different conditions to test the validity of the alterations to the model. We believe that such an effort is feasible and desirable.
Our proposed method for the statistical analysis of fMRI data seeks a compromise between validity, generality, simplicity and execution speed. The method is based on linear models with local AR(p) errors. The AR(p) model is fitted via the Yule-Walker equations with a simple bias correction that is similar to the first step in the Fisher scoring algorithm for finding ReML estimates. The resulting effects are then combined across runs in the same session, across sessions in the same subject, and across subjects within a population by a simple mixed effects model. The model is fitted by ReML using the EM algorithm after re-parameterization to reduce bias, at the expense of negative variance components. The residual degrees of freedom are boosted using a form of pooling by spatial smoothing. Activation is detected using Bonferroni, False Discovery Rate, and non-isotropic random field methods for local maxima and spatial extent. We briefly look at an alternative method based on conjunctions. Finally, we use a simple method to estimate and make inference about the delay of the hemodynamic response function at every voxel. We conclude with some suggestions for the optimal design of fMRI experiments.
Short Talk Abstract: Scientists use optical imaging - a functional neuroimaging technique with links to fMRI - to record stimulus-induced brain activity. Unfortunately, the measurable changes due to activation are overwhelmed by physiological fluctuations unrelated to the stimulus. Here I present early results from a method for removing this physiological noise to reveal the signal of interest, focusing first on the blood pressure. Two features distinguish this work: First, the data sets I consider are significantly larger than those usually analyzed: While a typical article in NeuroImage reports analysis of about 12 megabytes of raw optical imaging data per experiment, the 12-minute experiment I analyze here produced over 6 gigabytes of raw image data. Second, I present a scientifically appealing way of reducing the noise masking the signal of interest in the images: Where others use mathematical decompositions like PCA to separate signal from noise, I instead augment the images with simultaneously recorded physiological measurements. My procedure - multiple linear regression of the images on the blood pressure data - is simple statistically but challenging computationally: To compute my current model I perform almost 20 million regressions with over 20 thousand observations each.
Joint work with Bill Eddy and Seong-gi Kim.
Short Talk Abstract: The typical massively univariate approach to fMRI data analysis entails fitting 104-105 linear models, one at each voxel. In lieu of examining 100,000 sets of diagnostic plots, we propose a combination of statistical and graphical techniques to efficiently diagnose model fit on large datasets.
We create images of model summaries which assess standard linear modeling assumptions; for example, a cumulative periodogram statistic image is used to assess the degree of temporal dependence, and a Shapiro-Wilk statistic image to assess skew and outliers. We also create time series of image summaries; for example, global mean intensities and movement parameters. Together with a dynamic viewer that can jump from summary images to time series of residuals, or from summary time series to a series of residual images, we have systematic approach to model diagnosis that can identify subtle anomalies and important violations of model assumptions.
Poster Presentation Abstract: Technological advances in real-time fMRI including powerful data acquisition and analysis methods, and the use of higher field strength have greatly improved our ability to characterize changes in brain activation during the ongoing scan. We have developed TurboFIRE, an interactive real-time fMRI analysis tool interfaced to 1.5 and 4T MR scanners. Recent method development includes a novel spatial normalization method based on a look-up table approach with integrated Talairach Daemon database. This allows us to automatically label the location of activation clusters in terms of neuro-anatomical structures and to automatically select neuro-anatomically defined regions for ROIs analysis. Fitting and combination of multi-echo EPI data is performed on the fly to enhance BOLD sensitivity, General-Linear-Model Estimation, and simultaneous single- and multiple trial analysis are implemented, along with automated quantification of activation patterns. Applications range from quality control and neuroscience research using interactive paradigms, to clinical investigations.