Convergent methods for spatial and semantic image comparison
Abstract/Contents
- Abstract
- Mental illness encompasses a broad range of disorders that fill the lives of their sufferers with penetrating darkness. In the year 2014 it was estimated that 18.1% of U.S. adults experienced a mental illness in the past year [154], an alarming prevalence that has allowed mental illness to account for the highest percentage of disability adjusted life years globally [41]. Our current definition of mental illness stems from the Diagnostic and Statistical Manual (DSM), currently in its fifth edition [48], and while a behavioral description of disorder is highly useful for clinical diagnosis, unfortunately the disorder groups defined by the manual do not map onto biology [75]. In order to address this mental health crisis, researchers would need data-driven, biologically- oriented metrics associated with specific aberrant behavior and cognition. Toward this goal, the National Institute of Mental Health (NIMH) founded the Research Domain Criteria (RDoC) in 2009 [107]. This initiative would provide impetus for researchers to discover dimensions of genetic, neural, and behavioral features underlying neuropsychiatric disorder and mental illness. The RDoC initiative defines seven domains under which these different features can be studied, including cognition, social processes, arousal and regulatory systems, along with negative and positive valence systems [134]. This approach places an emphasis on linking core cognitive and developmental processes to behavior and symptoms, holding promise to redefine our definitions for mental illness to inspire better detection and prevention. It follows, then, that filling in this matrix calls for approaches to link cognitive processes to behavior. Our understanding of human cognition comes from approaches that can measure biological and cognitive variables during the live experience of one or more behaviors. Given these criteria, the primary avenue for understanding human cognition has come from non-invasive brain imaging. We rely on functional magnetic resonance imaging (fMRI) to non-invasively study the human brain across development, during the experience of different cognitive and behavioral tasks, and between different DSM-defined behavior groups. However, neuroimaging technology, supplemented by expertise in cognitive neuroscience, has been around for over two decades. Why, then, has cognitive neuroscience failed to characterize mental illness? A painful epiphany has been that more attention has not been given to studying what it means, computationally, to reproduce a result [141]. Our ability to synthesize the compendium of research in cognitive neuroscience relies primarily on meta analysis. Meta analysis across large datasets can provide validation of the existence of functional brain networks linked to cognitive processes, consensus about how different brain regions contribute to behavioral experience, and can lead to discovery of patterns that are not revealed among the noise of smaller datasets. Thus, while we have confidence that cognitive neuroscience has promise to better characterize mental illness, achieving this goal will require more standardization and best practices for synthesizing results, which come in the form of whole-brain statistical maps that reflect one or more cognitive processes of interest. Simply put, we must have informed ways to link cognitive processes and behavior to our measurement of it with functional brain imaging. One approach is to utilize the linguistic terms that describe these cognitive processes and the tasks that measure them (represented in an ontology), and then provide the infrastructure and means to annotate statistical brain maps with these terms. Using this annotation, we can then generate "encoding" models to predict entire statistical brain maps from terms alone. The responsibility to understand how similar one statistical brain map is to another, a procedure we might call "image comparison, " along with the task of properly mapping cognitive processes to our modality to measure them (fMRI) can be helped by Informatics. Informatics is a subset of science that focuses on the infrastructure and methodology of a scientific discipline. Identifying informed methods to assess the similarity of statistical brain maps, the most basic unit of output for neuroimaging scientists, is a logical first step toward helping cognitive neuroscience. This need to understand best practices and explore different strategies for image comparison motivates the work outlined in this thesis. My goal is to not only study methods, but to provide tools and infrastructure to compare brain images to drive meta analysis and relate cognitive processes to statistical brain maps. My approach consists of: 1) optimizing image comparison to assess the similarity of statistical brain maps based on spatial transformations, and 2) establishing novel semantic methods to drive image comparison based on cognitive processes. This second point is a valuable component of this work because it is currently not understood if the terminologies that are used in the field to represent different cognitive processes can be valuable in a classification framework. We define "valuable" as being comparable to a calculation of spatial similarity, a commonly done procedure. For each of these steps, I develop tools and infrastructure to immediately deploy relevant methods and empower researchers to compare their brain maps, and link fMRI to cognitive processes. Such an approach will link behaviors and cognition to specific neural pathways, allowing for a dimensional definition of these processes in the relevant domains of the RDoC matrix. A robust, automated derivation of a metric for fMRI image similarity based on these semantic annotations will be a novel contribution to this field. I will first show that there are optimal methods for spatial comparison of whole-brain statistical maps, the first fundamental step of any meta-analysis. I identify pairwise comparison methods that are in line with the goals of the researcher - optimal retrieval of maps from a particular experimental paradigm. This simple method of deriving a similarity score from a thresholded image, and only comparing spatial locations defined in both maps, results in 98.4% classification accuracy, and is robust across data derived from different individuals and a wide array of experimental paradigms. This work sets up a best practice for individual neuroimaging researchers to find "brain maps like mine, " and an informed basis for developers of tools that assess the reproducibility of neuroimaging results based on spatial information. Next, I present work to suggest that semantic information about the cognitive tasks that give rise to statistical brain maps can be useful in a classification framework. This finding demonstrates that semantic image comparison is useful for image classification, a simple function that is useful in many frameworks that require automated image labeling, and organization. Given this finding, I suggest that semantic classification should be further developed for statistical brain maps, as it offers a much computationally simpler comparison simply by way of the smaller data size. In this work, I will review automated and openly available methods for semantic classification of statistical brain maps, and demonstrate that such a comparison is comparable to the more commonly done calculation of spatial similarity. Adding a semantic dimension, meaning ascribing terms like "positive feedback" and "working memory" to an image comparison task is essential for several reasons. First, it brings a level of human interpretation to this framework. Second, ontology development is a continually developing science, and it is important to understand if effort in this domain is supported by the neuroimaging data that the ontologies describe. Finally, adding a semantic dimension to image comparison means that brain images derived from novel task structure can be interpreted, a procedure called "decoding" that holds promise to predict the cognitive concepts associated with any "unknown" brain map. As a supplement to this work, I include in this dissertation related work for standardization of experimental paradigms used to measure behavior associated with an fMRI protocol (Appendix A), along with a novel tool to visualize these functional networks (Appendix B). Finally, I will bring infrastructure and methods together, discussing best practices and additional tools that I have de- veloped during my graduate career to empower neuroscience researchers to assess the reproducibility of their work, and derive new reproducible products (Appendix C). This body of work, including tools and methods for image comparison and reproducible neuroscience, provides compelling exam- ple of both the strengths and challenges of the rising of informatics into the practice of reproducible neuroscience. Cognitive neuroscience will not move forward quickly enough as a science, bringing with it badly needed intervention for mental illness, if our researchers do not have better tools. The work outlined in this thesis focuses on that goal, and is the first of its type to span medical imaging expertise, machine learning and data visualization, as well as academic software development.
Description
| Type of resource | text |
|---|---|
| Form | electronic; electronic resource; remote |
| Extent | 1 online resource. |
| Publication date | 2016 |
| Issuance | monographic |
| Language | English |
Creators/Contributors
| Associated with | Sochat, Vanessa | |
|---|---|---|
| Associated with | Stanford University, Department of Biomedical Informatics. | |
| Primary advisor | Poldrack, Russell A | |
| Thesis advisor | Poldrack, Russell A | |
| Thesis advisor | O'Hara, Ruth (Ruth M.) | |
| Thesis advisor | Rubin, Daniel | |
| Advisor | O'Hara, Ruth (Ruth M.) | |
| Advisor | Rubin, Daniel |
Subjects
| Genre | Theses |
|---|
Bibliographic information
| Statement of responsibility | Vanessa Sochat. |
|---|---|
| Note | Submitted to the Department of Biomedical Informatics. |
| Thesis | Thesis (Ph.D.)--Stanford University, 2016. |
| Location | electronic resource |
Access conditions
- Copyright
- © 2016 by Vanessa Villamia Sochat
- License
- This work is licensed under a Creative Commons Attribution Non Commercial 3.0 Unported license (CC BY-NC).
Also listed in
Loading usage metrics...