How can we accumulate knowledge on brain functions? How can we leverage years of research in functional MRI to analyse finer-grained psychological constructs, and build a comprehensive model of the brain? Researchers usually rely on single studies to delineate brain regions recruited by mental processes. They relate their findings to previous works in an informal way by defining regions of interest from the literature. Meta-analysis approaches provide a more principled way to build upon the literature. This thesis investigates three ways to assemble knowledge using activation maps from a large amount of studies. First, we present an approach that uses jointly two similar fMRI experiments, to better condition an analysis from a statistical standpoint. We show that it is a valuable data-driven alternative to traditional regions of interest analyses, but fails to provide a systematic way to relate studies, and thus does not permit to integrate knowledge on a large scale. Because of the difficulty to associate multiple studies, we resort to using a single dataset sampling a large number of stimuli for our second contribution. This method estimates functional networks associated with functional profiles, where the functional networks are interacting brain regions and the functional profiles are a weighted set of cognitive descriptors. This work successfully yields known brain networks and automatically associates meaningful descriptions. Its limitations lie in the unsupervised nature of this method, which is more difficult to validate, and the use of a single dataset. It however brings the notion of cognitive labels, which is central to our last contribution. Our last contribution presents a method that learns functional atlases by combining several datasets. [Henson 2006] shows that forward inference, i.e. the probability of an activation given a cognitive process, is often not sufficient to conclude on the engagement of brain regions for a cognitive process. Conversely, [Poldrack 2006] describes reverse inference as the probability of a cognitive process given an activation, but warns of a logical fallacy in concluding on such inference from evoked activity. Avoiding this issue requires to perform reverse inference with a large coverage of the cognitive space. We present a framework that uses a "meta-design" to describe many different tasks with a common vocabulary, and use forward and reverse inference in conjunction to outline functional networks that are consistently represented across the studies. We use a predictive model for reverse inference, and perform prediction on unseen studies to guarantee that we do not learn studies' idiosyncrasies. This final contribution permits to learn functional atlases, i.e. functional networks associated with a cognitive concept. We explored different possibilities to jointly analyse multiple fMRI experiments. We have found that one of the main challenges is to be able to relate the experiments with one another. As a solution, we propose a common vocabulary to describe the tasks. [Henson 2006] advocates the use of forward and reverse inference in conjunction to associate cognitive functions to brain regions, which is only possible in the context of a large scale analysis to overcome the limitations of reverse inference. This framing of the problem therefore makes it possible to establish a large statistical model of the brain, and accumulate knowledge across functional neuroimaging studies.

XCEDE-DM: A neuroimaging extension to the W3C provenance data model

Large-scale functional MRI analysis to accumulate knowledge on brain functions

Reproducibility in human cognitive neuroimaging: a community-driven data sharing framework for provenance information integration and interoperability Boyce Nolan Nichols III Chair of the Supervisory Committee: Professor James F. Brinkley Biological Structure Access to primary data and the provenance of derived data are increasingly recognized as an essential aspect of reproducibility in biomedical research. While productive data sharing has become the norm in some biomedical communities, human brain imaging has lagged in open data and descriptions of provenance. The overarching goal of my dissertation was to identify barriers to neuroimaging data sharing and to develop a fundamentally new, granular data exchange standard that incorporates provenance as a primitive to document cognitive neuroimaging workflow. For my dissertation research, I led the development of the Neuroimaging Data Model (NIDM), an extension to the W3C PROV standard for the domain of human brain imaging. NIDM provides a language to communicate provenance by representing primary data, computational workflow, and derived data as bundles of linked Agents, Activities, and Entities. Similar to the way a sentence conveys a standalone thought, a bundle contains provenance statements that parsimoniously express the way a given piece of data was produced. To demonstrate a system that implements NIDM, I developed a modern, semantic Web application platform that provides neuroimaging workflow as a service and captures provenance statements as NIDM bundles. The course of this work necessitated interaction with an international community, which adopted

XCEDE-DM: A neuroimaging extension to the W3C provenance data model

Citations

Large-scale functional MRI analysis to accumulate knowledge on brain functions

Reproducibility in human cognitive neuroimaging: a community-driven data sharing framework for provenance information integration and interoperability

References

Large-scale functional MRI analysis to accumulate knowledge on brain functions

Reproducibility in human cognitive neuroimaging: a community-driven data sharing framework for provenance information integration and interoperability