Rozprawy doktorskie na temat „Computational neuroimaging”

This Thesis describes our work at the boundary between Computer Science and Cognitive (Neuro)Science. In particular, (1) we have worked on methodological improvements to clustering-based meta-analysis of neuroimaging data, which is a technique that allows to collectively assess, in a quantitative way, activation peaks from several functional imaging studies, in order to extract the most robust results in the cognitive domain of interest. Hierarchical clustering is often used in this context, yet it is prone to the problem of non-uniqueness of the solution: a different permutation of the same input data might result in a different clustering result. In this Thesis, we propose a new version of hierarchical clustering that solves this problem. We also show the results of a meta-analysis, carried out using this algorithm, aimed at identifying specific cerebral circuits involved in single word reading. Moreover, (2) we describe preliminary work on a new connectionist model of single word reading, named the two-component model because it postulates a cascaded information flow from a more cognitive component that computes a distributed internal representation for the input word, to an articulatory component that translates this code into the corresponding sequence of phonemes. Output production is started when the internal code, which evolves in time, reaches a sufficient degree of clarity; this mechanism has been advanced as a possible explanation for behavioral effects consistently reported in the literature on reading, with a specific focus on the so called serial effects. This model is here discussed in its strength and weaknesses. Finally, (3) we have turned to consider how features that are typical of human cognition can inform the design of improved artificial agents; here, we have focused on modelling concepts inspired by emotion theory. A model of emotional interaction between artificial agents, based on probabilistic finite state automata, is presented: in this model, agents have personalities and attitudes that can change through the course of interaction (e.g. by reinforcement learning) to achieve autonomous adaptation to the interaction partner. Markov chain properties are then applied to derive reliable predictions of the outcome of an interaction. Taken together, these works show how the interplay between Cognitive Science and Computer Science can be fruitful, both for advancing our knowledge of the human brain and for designing more and more intelligent artificial systems.

Depression and schizophrenia are common psychiatric disorders that can be disabling and chronic. This thesis aimed to further elucidate the underlying neural substrates using functional magnetic resonance imaging (fMRI) based studies. Hypothesized impairments in reinforcement learning in depression and schizophrenia were investigated, as were the neural correlates of abnormalities of social information processing in schizophrenia. Computational models of reinforcement learning are based on the concept of a 'prediction error' (PE, discrepancy between the expected and actual outcome) signal to update predictions of rewards and improve action selection. It has been argued that the firing of dopamine neurons encode a reward PE signal that mediates the learning of associations and the attribution of motivational salience to reward-related stimuli. Using model-based fMRI, the encoding of neural PE signals in patients with depression and schizophrenia were investigated. Consistent with hypotheses, patients exhibited different abnormalities in neural PE signals, with the degree of abnormality correlating with increased anhedonia/psychotic symptoms in depression/schizophrenia. These findings are consistent with the suggestion that a disruption in the encoding of PE signals contributes to anhedonia symptoms in depression by disrupting learning and the acquisition of salience of rewarding events. In schizophrenia, abnormal PE signals may contribute to psychosis by promoting aberrant perceptions and abnormal associations. In a different study, the neural responses to social exclusion in schizophrenia were investigated. Schizophrenia patients failed to modulate activity in the medial prefrontal cortex with the degree of exclusion, unlike controls. This highlights the neural substrates of putatively impaired social information processing in schizophrenia. Overall, these findings are consistent with proposals that psychiatric syndromes reflect different disorders of neural valuation. This perspective may help bridge the gap between the biological and phenomenological levels of understanding of depression and schizophrenia, hopefully contributing in the long term to the development of more effective treatments.

A statistical meta-analysis combines the results of several studies that address a set of related research hypotheses, thus increasing the power and reliability of the inference. Meta-analytic methods are over 50 years old and play an important role in science; pooling evidence from many trials to provide answers that any one trial would have insufficient samples to address. On the other hand, the number of neuroimaging studies is growing dramatically, with many of these publications containing conflicting results, or being based on only a small number of subjects. Hence there has been increasing interest in using meta-analysis methods to find consistent results for a specific functional task, or for predicting the results of a study that has not been performed directly. Current state of neuroimaging meta-analysis is limited to coordinate-based meta-analysis (CBMA), i.e., using only the coordinates of activation peaks that are reported by a group of studies, in order to "localize" the brain regions that respond to a certain type of stimulus. This class of meta-analysis suffers from a series of problems and hence cannot result in as accurate results as desired. In this research, we describe the problems that existing CBMA methods are suffering from and introduce a hierarchical mixed-effects image-based metaanalysis (IBMA) solution that incorporates the sufficient statistics (i.e., voxel-wise effect size and its associated uncertainty) from each study. In order to improve the statistical-inference stage of our proposed IBMA method, we introduce a nonparametric technique that is capable of adjusting such an inference for spatial nonstationarity. Given that in common practice, neuroimaging studies rarely provide the full image data, in an attempt to improve the existing CBMA techniques we introduce a fully automatic model-based approach that employs Gaussian-process regression (GPR) for estimating the meta-analytic statistic image from its corresponding sparse and noisy observations (i.e., the collected foci). To conclude, we introduce a new way to approach neuroimaging meta-analysis that enables the analysis to result in information such as “functional connectivity” and networks of the brain regions’ interactions, rather than just localizing the functions.

Recent advancements in miniaturized fluorescence microscopy have made it possible to investigate neuronal responses to external stimuli in awake behaving animals through the analysis of intra-cellular calcium signals. An ongoing challenge is deconvolving the noisy calcium signals to extract the spike trains, and understanding how this activity is affected by external stimuli and conditions. In this thesis, we aim to provide novel approaches to tackle various aspects of the analysis of calcium imaging data within a Bayesian framework. Following the standard methodology to the analysis of calcium imaging data based on a two-stage approach, we investigate efficient computational methods to link the output of the deconvolved fluorescence traces with the experimental conditions. In particular, we focus on the use of Poisson regression models to relate the number of detected spikes with several covariates. Motivated by this framework, but with a general impact in terms of application to other fields, we develop an efficient Metropolis-Hastings and importance sampling algorithm to simulate from the posterior distribution of the parameters of Poisson log-linear models under conditional Gaussian priors, with superior performance with respect to the state-of-the-art alternatives. Motivated by the lack of clear uncertainty quantification resulting from the use of a two-stage approach, and the impossibility to borrow information between the two stages, we focus on the analysis of individual neurons, and develop a coherent mixture model that allows for estimation of spiking activity and, simultaneously, reconstructing the distributions of the calcium transient spikes' amplitudes under different experimental conditions. More specifically, our modeling framework leverages two nested layers of random discrete mixture priors to borrow information between experiments and discover similarities in the distributional patterns of the neuronal response to different stimuli. Finally, we move to the multivariate analysis of populations of neurons. Here the interest is not only to detect and analyze the spiking activity but also to investigate the existence of groups of co-activating neurons. Estimation of such groups is a challenging problem due to the need to deconvolve the calcium traces and then cluster the resulting latent binary time series of activity. We describe a nonparametric mixture model that allows for simultaneous deconvolution and clustering of time series based on common patterns of activity. The model makes use of a latent continuous process for the spike probabilities to identify groups of co-activating cells. Neurons' dependence is taken into account by informing the mixture weights with their spatial location, following the common neuroscience assumption that neighboring neurons often activate together.

Adaptation is a fundamental property of the nervous system that underlies the maintenance of successful actions through flexible reconfiguration of sensorimotor processing. The primary aims of this thesis are 1) to investigate the computational and neural underpinnings of sensorimotor memory formation during prism adaptation (PA) in humans, and 2) how they interact with anodal transcranial direct current stimulation (a-tDCS) of the primary motor cortex (M1), in order to 3) improve efficacy of prism therapy for post-stroke spatial neglect. In chapter 4, we modify an influential state-space model of adaptation in order to characterize the contribution of short and long memory timescales to motor behaviour as sensorimotor after-effects (AEs) develop during PA. This enables us, in the multimodal 7 Tesla MRI experiment reported in chapter 5, to demonstrate that the level of M1 excitation:inhibition causally sets the relative contribution of long versus short memory timescales during PA, thus determining behavioural persistence of the AE at retention in young healthy adults. This finding offers a bridge between different levels of investigation by providing a biologically plausible neuro-computational model of how sensorimotor memories are formed and enhanced by a-tDCS. In chapter 6, we use the ageing motor system as a model of reduced GABAergic inhibition and show that the age-related decrease in M1 GABA explains why older adults demonstrate more persistent prism AEs. Taken together, these data indicate that the reduction in M1 GABAergic inhibition via excitatory a-tDCS during PA has the potential to enhance persistence of adaptation memory in both young and older adults. Informed by these results, we subsequently ask whether standard (multi-session) PA therapy combined with left M1 a-tDCS translates to greater and/or longer-lasting clinical improvements in post-stroke spatial neglect patients. In chapter 7, we compare the multimodal neuroimaging data of six neglect patients to normative data of age-matched controls. We show that in all patients, the lesion interrupted long-range frontoparietal connections, and we provide direct evidence for a pathological left dominance of activity within the lateral occipital cortex during deployment of bilateral visuospatial attention. In chapter 8, we present the behavioural performance of these patients throughout the two phases of the clinical study (i.e. before and after either PA + real M1 a-tDCS or PA + sham M1 atDCS). There was no clear effect of a-tDCS on the therapeutic effect of PA in these patients. The results of the studies presented in this thesis provide a novel insight into the neurocomputational mechanisms of sensorimotor memory formation and its modulation by a-tDCS in the healthy brain. Further investigation of how these mechanisms relate to therapeutic improvements following PA in certain neglect patients is needed.

L'étude des relations entre plusieurs ensembles de variables mesurées sur un même groupe d'individus est un défi majeur en statistique. La littérature fait référence à ce paradigme sous plusieurs termes : "analyse de données multimodales", "intégration de données", "fusion de données" ou encore "analyse de données multibloc". Ce type de problématique se retrouve dans des domaines aussi variés que la biologie, la chimie, l'analyse multi-capteurs, le marketing, la recherche agro-alimentaire, où l'objectif commun est d'identifier les variables de chaque bloc intervenant dans les intéractions entre blocs. Par ailleurs, il est possible que chaque bloc soit composé d'un très grand nombre de variables (~1M), nécessitant le calcul de milliards d'associations. L'élaboration d'un cadre statistique épousant la complexité et l'hétérogénéité des données est donc primordial pour mener une analyse pertinente.Le développement de méthodes d'analyse de données hétérogènes, potentiellement de grande dimension, est au coeur de ce travail. Ces développements se basent sur l'Analyse Canonique Généralisée Régularisée (RGCCA), un cadre général pour l'analyse de données multiblocs. Le coeur algorithmique de RGCCA se résume à un unique "update", répété jusqu'à convergence. Si cet update possède certaines "bonnes" propriétés, la convergence globale de l'algorithme est garantie. Au cours de ces travaux, le cadre algorithmique de RGCCA a été étendu dans plusieurs directions :(i) Du séquentiel au global. Plutôt que d'extraire de chaque bloc les composantes de manière séquentielle, un problème d'optimisation globale permettant de construire ces composantes simultanément a été proposé.(ii) De la matrice au tenseur. L'Analyse Canonique Généralisée Multivoie (MGCCA) étend RGCCA à l'analyse conjointe d'un ensemble de tenseurs. Des versions séquentielle et globale de MGCCA ont été proposées. La convergence globale de ces algorithmes est montrée.(iii) De la parcimonie à la parcimonie structurée. Le coeur de l'algorithme d'Analyse Canonique Généralisée Parcimonieuse (SGCCA) a été amélioré en fournissant un algorithme à convergence globale beaucoup plus rapide. Des contraintes de parcimonie structurée ont également été ajoutées à SGCCA.Dans une seconde partie, l'analyse de plusieurs jeux de données est menée à l'aide de ces nouvelles méthodes. La polyvalence des ces outils est démontrée sur (i) deux études en imagerie-génétique, (ii) deux études en électroencéphalographie ainsi (iii) qu'une étude en microscopie Raman. L'accent est mis sur l'interprétation des résultats facilitée par la prise en compte des structures multiblocs, tensorielles et/ou parcimonieuses
A challenging problem in multivariate statistics is to study relationships between several sets of variables measured on the same set of individuals. In the literature, this paradigm can be stated under several names as “learning from multimodal data”, “data integration”, “data fusion” or “multiblock data analysis”. Typical examples are found in a large variety of fields such as biology, chemistry, sensory analysis, marketing, food research, where the common general objective is to identify variables of each block that are active in the relationships with other blocks. Moreover, each block can be composed of a high number of measurements (~1M), which involves the computation of billion(s) of associations. A successful investigation of such a dataset requires developing a computational and statistical framework that fits both the peculiar structure of the data as well as its heterogeneous nature.The development of multivariate statistical methods constitutes the core of this work. All these developments find their foundations on Regularized Generalized Canonical Correlation Analysis (RGCCA), a flexible framework for multiblock data analysis that grasps in a single optimization problem many well known multiblock methods. The RGCCA algorithm consists in a single yet very simple update repeated until convergence. If this update is gifted with certain conditions, the global convergence of the procedure is guaranteed. Throughout this work, the optimization framework of RGCCA has been extended in several directions:(i) From sequential to global. We extend RGCCA from a sequential procedure to a global one by extracting all the block components simultaneously with a single optimization problem.(ii) From matrix to higher order tensors. Multiway Generalized Canonical Correlation Analysis (MGCCA) has been proposed as an extension of RGCCA to higher order tensors. Sequential and global strategies have been designed for extracting several components per block. The different variants of the MGCCA algorithm are globally convergent under mild conditions.(iii) From sparsity to structured sparsity. The core of the Sparse Generalized Canonical Correlation Analysis (SGCCA) algorithm has been improved. It provides a much faster globally convergent algorithm. SGCCA has been extended to handle structured sparse penalties.In the second part, the versatility and usefulness of the proposed methods have been investigated on various studies: (i) two imaging-genetic studies, (ii) two Electroencephalography studies and (iii) one Raman Microscopy study. For these analyses, the focus is made on the interpretation of the results eased by considering explicitly the multiblock, tensor and sparse structures

Neuropsychological traits affect both the brain and behaviour and are responsible for a large proportion of worldwide disability. This PhD thesis employs computational, statistical and genetic approaches to identify and understand the genetic and environmental influences on a wide range of psychiatric, neurological and cognitive disorders. The work presented in this thesis details novel findings on several fronts including new genetic marker discovery, using genetics to predict an individual’s disease risk, and disentangling pertinent risk factors that affect cognitive and mental health. This insight is an important step towards developing more effective treatments and intervention strategies.

A new research direction in the neuroimaging discipline, so called imaging genetic, has emerged recently concerns describing individual differences in imaging phenotypes using genetic and environmental factors. The large number of voxel- and vertex-wise measurements in imaging genetics studies present a challenge both in terms of computational intensity and the need to account for elevated false positive risk because of the multiple testing problem. There is a gap in existing tools, as standard neuroimaging software cannot perform essential genetic analyses including heritability and association estimations and testings, and yet standard quantitative genetics tools cannot provide essential neuroimaging inferences, like family-wise error corrected voxel- wise or cluster-wise P-values. Moreover, available genetic tools rely on P-values that can be inaccurate with usual parametric inference methods. In this thesis computationally efficient linear mixed effect model for voxel-wise genetic analyses of high-dimensional imaging phenotypes are developed. Specifically, fast estimation and inference procedures for heritability and association analyses are introduced using orthogonal transformations that dramatically simplify the likelihood and restricted likelihood functions of mixed effect model. We review the family of score, likelihood ratio and Wald tests and propose novel inference methods for fixed and random effect terms in the mixed effect models. To address problems with inaccuracies with the standard results used to find P-values, we propose different permutation schemes to allow semi-parametric inference (parametric likelihood-based estimation, non-parametric sampling distribution). In total, we evaluate different significance tests for heritability and association, with either asymptotic parametric or permutation-based P-value computations. We identify a number of tests that are both computationally efficient and powerful, making them ideal candidates for heritability and genome-wide association studies in the massive data setting.

Real-world decision making is challenging due, in part, to changes in the underlying reward structure: the best option last week may be less rewarding today. Determining the best response is even more challenging when feedback validity is low. Presented here are the results of two experiments designed to determine the degree to which midbrain reward processing is responsible for detecting reward contingency changes when feedback validity is low. These results suggest that while midbrain reward systems may be involved in detecting unexpected uncertainty in non-stationary environments, other systems are likely involved when feedback validity is low – namely, the locus-coeruleus-norepinephrine system. Finally, a computational model that combines these systems is described and tested. Taken together, these results downplay the role of the midbrain reward system when feedback validity is low, and highlight the importance of the locus-coeruleus-norepinephrine system in detecting reward contingency changes.

This dissertation develops computational models to address outstanding problems in the domain of expectation-related cognitive processes and their neuroimaging markers in functional MRI or EEG. The new models reveal a way to unite diverse phenomena within a common framework focused on dynamic neural encoding shifts, which can arise from robust interactive effects of M-currents and chloride currents in pyramidal neurons. By specifying efficient, biologically realistic circuits that achieve predictive coding (e.g., Friston, 2005), these models bridge among neuronal biophysics, systems neuroscience, and theories of cognition. Chapter one surveys data types and neural processes to be examined, and outlines the Dynamically Labeled Predictive Coding (DLPC) framework developed during the research. Chapter two models hippocampal prediction and mismatch, using the DLPC framework. Chapter three presents extensions to the model that allow its application for modeling neocortical EEG genesis. Simulations of this extended model illustrate how dynamic encoding shifts can produce Mismatch Negativity (MMN) phenomena, including pharmacological effects on MMN reported for humans or animals. Chapters four and five describe new modeling studies of possible neural bases for alpha-induced information suppression, a phenomenon associated with active ignoring of stimuli. Two models explore the hypothesis that in simple rate-based circuits, information suppression might be a robust effect of neural saturation states arising near peaks of resonant alpha oscillations. A new proposal is also introduced for how the basal ganglia may control onset and offset of alpha-induced information suppression. Although these rate models could reproduce many experimental findings, they fell short of reproducing a key electrophysiological finding: phase-dependent reduction in spiking activity correlated with power in the alpha frequency band. Therefore, chapter five also specifies how a DLPC model, adapted from the neocortical model developed in chapter three, can provide an expectation-based model of alpha-induced information suppression that exhibits phase-dependent spike reduction during alpha-band oscillations. The model thus can explain experimental findings that were not reproduced by the rate models. The final chapter summarizes main theses, results, and basic research implications, then suggests future directions, including expanded models of neocortical mismatch, applications to artificial neural networks, and the introduction of reward circuitry.

Intact structural connectivity among brain regions is critical to cognition. Structural connectivity forms the substratum for information flow between brain regions, and its plasticity is a hallmark of learning in the brain. Moreover, structural connectivity markers constitute a heritable phenotype. Investigating neuroanatomical connectivity in the human brain is, therefore, critical not only for uncovering the neural underpinnings of behavior but also for understanding connectomic bases of neurodevelopmental and neurodegenerative disorders, such as autism and Alzheimer’s Disease. Diffusion magnetic resonance imaging (dMRI) and tractography are among the only techniques, at present, that enable estimation of anatomical connectivity in the human brain, in-vivo. By tracking the anisotropic diffusion of water molecules in white matter, dMRI and tractography enable post hoc reconstruction of contiguous fascicles between distal brain regions. How accurately can dMRI and tractography track these connections to match ground-truth in the brain? Are structural connections between specific pairs of brain regions informative about subjects’ cognitive capacities, like attention? Could changes in these connections indicate mechanisms of cognitive decline, both in healthy and pathologically aging populations? In this thesis, I report results from three studies, each of which addresses one of these key questions. In the first study, I explored how the midbrain contributes to attention, by combining model-based analysis of behavior with dMRI-tractography. Specifically, I examined the role of the superior colliculus (SC), a vertebrate midbrain structure, in attention. Does the SC control perceptual sensitivity to attended information, does it enable biasing choices toward attended information, or both? I mapped structural connections of the human SC with neocortical regions and found that the strengths of these connections correlated with, and were strongly predictive of, individuals’ choice bias, but not sensitivity. Taken together with previous studies, these results indicate that the human SC may play an evolutionarily conserved role in controlling choice bias during visual attention. In the second study, I developed a novel approach, implemented on GPUs, for pruning structural connectomes, at scale. First, I identified key limitations of a state-of-the-art connectome pruning technique, Linear Fascicle Evaluation (LiFE), and introduced a GPU-based implementation that achieves >100x speedups over conventional CPU-based implementations. Leveraging these speedups, I advanced LiFE’s algorithm by imposing a regularization constraint on estimated fiber weights. This regularized, accelerated, LiFE algorithm (“ReAl-LiFE”) estimates sparser connectomes that also provide more accurate fits to the underlying diffusion signal, and enables rapid and accurate connectome evaluation at scale. In the third study, I demonstrated several real-world applications of the ReAl-LiFE technique for analysis of large datasets. First, I showed that structural connectivity estimated with ReAl-LiFE predicts behavioral scores across a range of cognitive tasks in a cohort with 200 healthy human volunteers from the Human Connectome Project database. Moreover, ReAl-LiFE pruned connection weights provided a more reliable marker for structural connectivity strength than the number of fibers in the unpruned connectome. Second, ReAl-LiFE connection weights effectively predicted both chronological age, as well as age-related decline in cognitive factor scores in a cohort of 101 healthy, aged volunteers whose data were acquired as part of the Tata longitudinal study on aging at IISc. Finally, analyzing nearly 100 dMRI scans from the ADNI database, I showed that ReAl-LiFE outperformed competing approaches in terms of its accuracy with classifying patients with Alzheimer’s Dementia from healthy, age-matched controls, based on cortico-hippocampal connection weights. In summary, these findings show that diffusion MRI and tractography can serve as powerful tools for addressing key questions regarding brain-behavior relationships. In this thesis, I developed a technique to reliably estimate structural connectivity between distal brain regions, identified the role of subcortical structural connections in attention, and showed that cortical connectivity can be used to predict behavioral scores and cognitive decline. Broadly, these results will be relevant for understanding the connectomic basis of various cognitive processes, like attention, in healthy populations, and its dysfunction in diseased patients.