Poster Session B, Sunday, March 24, 8:00 – 10:00 am, Pacific Concourse
DeLINEATE: A deep learning toolbox for neuroimaging data analysis
Karl Kuntzelman1, Jacob M. Williams1, Ashok Samal1, Prahalada K. Rao1, Matthew R. Johnson1; 1University of Nebraska-Lincoln
Brain decoding, the use of machine learning techniques to identify cognitive states from neuroimaging data, has rapidly grown in popularity since the advent of multivariate pattern analysis (MVPA). The relatively simple linear algorithms that underlie most MVPA, however, are computationally inadequate for large and complex datasets, and in at least some domains (e.g., image classification) easily outperformed by deep neural networks (DNNs). Consequently there has been a recent surge of interest in approaches to brain decoding that capitalize on these “deep learning” techniques. Still, adoption of DNNs for MVPA has been slowed by the greater complexity of DNN architectures, the programming expertise required to use current DNN tools, and the lack of built-in support for specific methodological requirements of the neuroimaging community. To address these issues, we have created an open-source Python toolbox called Deep Learning In Neuroimaging: Exploration, Analysis, Tools, and Education (DeLINEATE). Notable features include support for several cross-validation and rescaling schemes; a text-based job description file format that requires no Python coding knowledge from users; and backend support for PyMVPA, enabling comparisons to traditional MVPA techniques. Across multiple datasets analyzed with DNNs, we demonstrate advantages in computational speed and/or classification performance relative to traditional MVPA techniques. DNNs can also implement novel forms of MVPA and tackle research questions that were unavailable to older methods. The current release is available on the project website (http://delineate.it) and all project code is hosted on Bitbucket (https://bitbucket.org/delineate/delineate/src/master/). Development is ongoing and we invite feature requests from the community.
Topic Area: METHODS: Neuroimaging