You are here
Magnetic resonance diffusion imaging and computational
tractography are the only technologies that enable neuroscientists to
measure white matter in the living human brain. In the decade since
their development, these technologies revolutionized our understanding
of the importance of the human white-matter for health and disease.
There are good reasons to make these measurements in human. The human
brain (1400 g) is 15 times the volume of the rhesus monkey (90 g), 700
times the volume of the rat (2 g) and 2,300 times the volume of the
mouse brain (0.6 g). The human brain comprises of functionally
specific clusters of maps communicating via an extensive network of
long-range, myelinated, axonal projects. The size of the human brain
imposes significant challenges for communicating across different
Prior to these technologies, the white matter was thought of as a
passive cabling system. But modern measurements show that white matter
axons and glia respond to experience and that the tissue properties of
the white matter are transformed during development and following
training. The white matter pathways comprise a set of active wires and
the responses and properties of these wires predict human cognitive
and emotional abilities in health and disease. We can now predict
confidently that to crack the neural code in mapping the human brain,
neuroscientists will have to develop an account of the connections and
tissue properties of these active wires. Whereas there are many
impressive findings, it is widely agreed that there is an urgent need
to keep developing and improving tractography methods. The need for a
systematic approach to tractography validation and for a framework to
perform statistical model testing in individual brains has been
claimed (Pestilli Nature: Scientific Data 2015).
First, I will present new framework for performing tractography
evaluation and statistical inferecne on the network of brain
connections (Pestilli et al., Nature Methods 2014). Second, I will
introduce recent advances in methods for mapping human connectomes in
living individuals (Takemura et al., PLoS computational Biology 2016).
These new methods improve current techniques in fundamental ways and
can be applied to any type of diffusion data.Finally, I will briefly
show that by using the methods we were able to identify a major
white-matter pathways previously unreported in the human brain. Such
as sensory-motor integration (Takemura et al., Cerebral Cortex 2015;
Yeatman et al., PNAS 2014), object-perception (Gomez et al., Neuron
2015) and decision making related pathways (Leong et al., Neuron
Neuroimaging methods provide a powerful tool to investigate brain connectivity, and have been widely applied to study the mutual relationship between structural and functional connections between brain regions in healthy subjects and in patients. Several studies have demonstrated that structurally connected cortical regions in the adult, healthy brain exhibit stronger and more consistent functional connectivity than structurally unconnected regions. However, the picture that emerges from studies in patients affected by brain disease and neurological conditions is more complex. A striking example is that of subjects with Agenesis of the Corpus Callosum (ACC), a congenital condition whereby the main bundle of white matter fibers connecting the two cerebral hemispheres does not form during brain development. Recent studies in ACC subjects, whose structural connectivity is dramatically impaired, have shown intact bilateral functional connectivity patterns, thus challenging the view that structural and functional connectivity are straightforwardly related. The underpinnings of the abnormal relationship between structural and functional connectivity in this and other brain disorders are unknown, and their investigation may provide insight into the underlying pathophysiological mechanisms, and possibly into compensatory mechanisms that may promote recovery of functional connectivity in case of congenital or acquired white matter loss.
The ability to quantitatively compare structural and functional connectivity would be important to assess the relationship between these two measures of interregional connections in the healthy brain and in diseased states. While selected connectivity paths can be assessed with seed-region-based methods, there is no generally established hypothesis-free approach to measuring global differences in different connectivity measures. Graph-theoretical analysis is attracting increasing attention as a general and powerful framework to analyze brain connectivity networks, and may provide a rigorous theoretical basis to study connectivity and disconnectivity in the brain. Here, I shall describe recent developments in complex network theory and its application to the study of topological aspects of brain connectivity, including the relation between structural and functional connectivity in models of brain disease. I will also discuss open problems in brain connectivity analysis that may be amenable to solution via graph-theoretical methods.
Many of the classification algorithms developed in the machine learning literature, including the support vector machine and boosting, can be viewed as methods that minimize a convex surrogate of the 0-1 loss. The convexity makes these algorithms computationally efficient. The use of a surrogate, however, has statistical consequences that must be balanced against the computational virtues of convexity. One property that establishes the optimality (under assumptions) for a classifier obtained by a minimization procedure such as support vector machine is that of Fisher consistency. In this talk I will describe first the notion of Fisher consistency on this setting and then describe the main results that have been emerged in recent years. In particular, I will review results for binary, multiclass classification and ranking algorithms. I will finish by stating novel results obtained for ordinal regression methods.
How do we make sense of words? What are the neural mechanisms by which a simple arrangement of strokes may trigger rich and elaborated mental representations? Meaning is a multidimensional representation that includes both abstract features (e.g. taxonomy), as well as perceptual features (e.g. prototypical size, color) of the objects referred to by words. What are the neural systems coding for these different conceptual and perceptual dimensions of word meaning? How do they articulate in time to give rise to the impression of a unitary representation of meaning? Our first study suggests the existence of a postero-to-anterior gradient of information coding along the ventral pathway: from purely perceptual (e.g. prototypical size) to conceptual (e.g. semantic category). Capitalizing from what we have learned, we conducted a combined MEG/fMRI experiment in order to extend our investigation in two main directions: (1) add another perceptual dimension, studying the patterns associated not only with visual properties but also auditory ones; (2) add information on the temporal dynamics thanks to MEG high time resolution. The reactivation of visual features in primary visual areas corresponds to a similar reactivation of auditory features in primary auditory areas? Do we first activate the more abstract semantic features in anterior temporal lobes, and only after the perceptual features in perceptual regions, or is the temporal order of activation the reverse? Alternatively, do these different representations get activated entirely in parallel? Analyzing the data with multivariate methods (i.e. classifying and correlating the different patterns of activation recorded) we will try to shed light on when and where a given stimulus dimension is coded and how brain activity changes in time and space to accomplish the goal of transforming meaningless symbols into unitary meaningful concepts.
Data analysis for functional Near InfraRed Spectroscopy (fNIRS) has evolved in the last twenty years, from well-known saturation-like analysis to modern Diffuse Optical Tomography (DOT). Moreover, anatomically informed fNIRS can now complement the classical fNIRS analysis and help the cortical localization of the brain signals. In this presentation, the physical principles of fNIRS and fNIRS instrumentation will be first discussed. Then, we will fly over data analysis of functional information, in particular discussing time analysis, depth analysis and space analysis of fNIRS data.
Real-time fMRI has become a valuable tool in neuroscience. RtfMRI creates "added value" as compared with offline analysis in several ways: we can see something happen in real time (BCI), we can manipulate cognitive states online (neurofeedback and brain-state dependent stimulation). We could also run tests in real-time (like localizer), test hypotheses about cognitive states or subject's parameters (perceptual threshold or bias). That could give us an advantage in time: in one session we could have results that would otherwise require multiple sessions. The key feature that would allow us to benefit from analysis in real-time is the possibility to adapt stimulation protocols online. However, if we turn to the state of the art in current rtfMRI experiments, we can see that there is a notable gap in this area. In my presentation, I would like to discuss the implementation of an adaptive protocol for our ATTEND real-time experiment with Adaptive Design Optimization framework (Daunizeau et al. ) and how this setup could be generalized to run other types of experiments with rtfMRI.
Diffusion magnetic resonance imaging (dMRI) allows to study brain connectivity by reconstructing the 3D pathways of axons within the white matter of the brain as a set of streamlines called tractography. Due to the heterogeneous nature of the brain, tractographies of different brain cannot be represented in a common space. Therefore for group-based analysis of different tractographies, alignment process is require. The most commonly used technique to do alignment is registration done by calculating the transformation from anatomical MR images. In contrast to registration techniques, we exploit the relationship between streamlines to find the correspondence between different tractographies without any transformation. In this perspective, we reform the problem of alignment as a problem of mapping. We formulate the mapping problem as variant of the graph matching problem. Here, the biggest challenge is to find the appropriate optimization technique with low computational complexity for tractography mapping.
A few recent studies (Reddy et al. 2010, Cichy et al 2011) have shown that there is common neural basis between perception and imagery. In these studies it was shown that it was not only possible to decode object category of perceived or imagined objects from object selective voxels in the brain, but decoding worked also in a cross modal manner. That is, the classifier trained on the activity pattern in object selective cortex during perception trials could reliably decode the object category of the imagery trials and vice a versa. However, it was shown that decoding of perceived objects using as training data imagery trials works somewhat better than in the other direction.
I presented the results of the analysis of the experiment data where subjects in half of the trials viewed images of people and cars and in the other half of the trials imagined people or cars according to the cue. In the decoding analysis, I replicated the results of the previous studies on cross-modal decoding of perception and imagery using he data of 10 subjects. Below are the accuracies obtained with 4-fold cross-validation and averaged over the number of subjects. First, it was possible to decode the object category with reliable accuracy within modality: it was around 90% in low visual areas and above 70% in ITG and fusiform cortex for perception and around 64% in object selective cortex (OSC) for imagery. Next, for cross-modal decoding I had 64% accuracies in OSC when the classifier was trained on imagery data and tested on perception data and 60% when the classifier was trained on perception data and tested on imagery trials. Apart from replicating the previous results, the goals of the present study are: 1) to use perception and imagery data for cross-modal decoding of visual search preparation data and 2) for setting up a real time fMRI experiment where the object identity during visual search preparation could be decoded online.