Abstract
This presentation argues that many currently popular strategies for analyzing BOLD functional imaging data should be replaced by methods that respect the fine structure of the human brain. It is well known that the human brain has ~1011 neurons, ~1011 astrocytes, and ~1014 synapses. In cadaver brain, the cortex can be parcellated into more than 50 areas of recognizably different cyto and myeloarchitecture. Each cortical area is connected by axons to at least 10 other areas, on average, and each cortical area is connected to several thalamic nuclei, basal ganglia, and specific cerebellar areas. Most white matter voxels contain crossing fibres. Cortex is always less than 4 mm thick, and many cortical areas have sharp boundaries: brain activity is generally not smooth. Significant details defining cortical areas are on a scale < 300 micrometres, and the vascular consequences of changes in neural activity are confined to the cortex and pial vessels.
Despite these facts, the conventional explicit and/or implicit assumptions that underlie the use of SPM, FSL, and other packages designed for fMRI analysis can be stated as follows: (i) Statistical inference using fMRI data is only possible if the image is smoothed by a kernel three times the size of the acquired voxels. (ii) Neural Mass Modelling is valid everywhere in the brain, at a scale of > 8 mm. (iii) Brain architecture is conserved across brains at a scale of 8 mm, and is defined by a probabilistic brain atlas valid for all adult brains. (iv) The importance of a given brain area for a specific task is indicated only by a positive mean activity, on a scale of 8 mm. (v) Activity in any given brain voxel is statistically independent from any other brain voxel. (vi) Grey matter, white matter and CSF are considered to be equally likely, as possible locations of changed brain activity. (vii) If a given voxel’s task-related variance is smaller than a statistical threshold of p < 0.05, after correction for multiple comparisons, it can play no role in the task performance. (viii) The centroid of a thresholded activated area completely defines its localization. (ix) Typical networks of active brain areas contain no more than 3-4 point-like nodes, representing spatially extended regions.
For any scientist who has taken the trouble to study neuroanatomy and neurophysiology, and to track developments in image statistical analysis, these assumptions are all obviously unrealistic. Neural representations are often on a distance scale of 1 mm or less, and BOLD signal is overwhelmingly confined to grey matter and the pial surface. Brain activity is inherently multivariate, and it is not smooth. Decreases in neural activity are equally as important as increases for the current powerful methods of pattern classification. It is grotesque to simplify brain network activity to as few as 3-4 nodes.
The key step in such conventional analyses of brain function, which has led to these strange assumptions, is spatial smoothing. The motivations for spatial smoothing are these: a) to increase SNR by averaging over adjacent voxels, b) to allow statistical inference of significance using the theory of Gaussian Random Fields, and c) to enable averaging of functional imaging data across groups of subjects, after each subject’s brain images have been suitably normalized into a standard template brain volume. The most unfortunate downside of smoothing is that it inextricably conflates the spatial extent and amplitude of brain activations. Each of these is a highly important variable in the understanding of grey matter operations. What is also lost is spatial precision, so that it becomes no longer possible to state, for instance, on which bank of a sulcus the activation actually lies.
Current progress in 7T MRI of human brain provide structural images with intracortical detail at 0.3 mm isotropic resolution, functional images with 0.65 mm isotropic resolution, and connectivity tractograms with 0.8 mm isotropic resolution. In such state-of-the-art data, the functional contrast-to-noise of the BOLD images is sufficiently high that averaging by smoothing is unnecessary. Image smoothness is not required for valid statistical thresholding using the False Discovery Rate (FDR) method. Finally, smoothing is no longer required for averaging across subjects, when structural MRI can provide a native map of the areal boundaries for each subject’s brain. This removes the guesswork implicit in probabilistic parcellations of grey matter, and enables realistic ROI’s to be outlined in each brain, resulting in a dramatic increase in experimental power and much greater spatial precision.
An example of a fresh approach to functional MRI analysis is presented, in the form of a BOLD fMRI study at 7T (Siemens Magnetom) of the cortical layer dependence of motor cortex activity, in the three conditions of finger tapping, finger movement without fingertip contact, and motor imagery of finger tapping. Using a BOLD isotropic voxel size of 0.75 mm, activity could be discriminated in each of four automatically computed contours of motor cortex, the third deepest of which corresponds to cortical output layer V. Because the hand area of motor cortex can be easily delineated in human brain, using the landmarks of the ´hand knob` in the central sulcus and the considerably reduced T1 found in the heavily myelinated motor cortex, averaging across the nine subjects studied could be performed using unsmoothed data, by regions of interest defined in each subject as contours of grey matter on the hand knob. The grand averaged results showed that activity in cortical layer V during the motor imagery condition was highly significantly reduced, as compared with the active motor conditions. This is consistent with the absence of motor output in the imagery condition. Such an approach can provide evidence for the causal role of specific cortical areas in the performance of many types of brain task, a capability which has previously been unavailable in noninvasive brain function studies, despite the unsuccessful attempts invoking Granger Causality and Dynamic Causal Modelling.
Coupling such experimental designs with high resolution crossing-fibre tractography is likely to lead to a far more precise understanding of how brain networks function in task performance. Current assumptions driven by spatial smoothing clearly lead to a sterile impasse. Only when the living brain’s fine intracortical details are available can more fruitful simplifying assumptions be made, to enable modelling powerful enough to allow testable predictions. Brain imaging at 7T appears to be capable of providing such details.
