BRAINSTORM Towards Clinically and Scientifically Useful NeuroImaging Analytics
Joshua Vogelstein (Johns Hopkins University), Sharad Sikka (Nathan Klein Institute), Brian Cheung (Nathan Klein Institute), Ranjit Khanuja (Child Mind Institute), Qinyang Li (Child Mind Institute), Yan Chao-Gan (Child Mind Institute), Carey Priebe (Johns Hopkins University), Vince Calhoun (Mind Research Network), R. Jacob Vogelstein (Johns Hopkins University), Michael Milham (Child Mind Institute), Randal Burns (Johns Hopkins University)
We desire to transform clinical psychiatric practice to take advantage of the vast technological strides in contemporary neuroimaging. We propose three complementary steps will help facilitate this transformation. First, the construction of a computing platform to store and process large datasets. Second, methods to calibrate measurements across individuals and instruments. Third, tools to convert such measurements into clinically useful analytics. We are developing BRAINSTORM (Fig. 1) to address these three concerns.
First, a high-performance compute cluster and associated scientific database, called "BrainCloud", for storing, managing, and efficiently querying both multi-modal neuroimaging and rich phenotypic data. BrainCloud will be seeded with data already available from the International NeuroImaging Data Initiative  as well as the Mind Research Network . Moreover, BrainCloud will include a simple one-click upload interface so that additional research and clinical facilities can contribute to the growing data corpus.
Second, a robust pipeline optimized to pre-process multimodal image data to infer multi-modal attributed connectomes (MACs). We are developing a highly configurable pipeline  that enables us to search for an optimal representation of data for subsequent inference via non-parametric reliabilities estimates.
Third, streaming decision theoretic manifold learning algorithms  that yield clinically useful outputs, as well as provide insight into brain/behavior relationships. To date, most statistical and machine learning algorithms natively operate on vector valued data; but our data are far more complex: responses to psychological instruments and multimodal images. We are developing complementary tools that natively operate on non-Euclidean data and "stream", meaning that they continue to learn as new data becomes available.
Schematic Illustration of BrainStorm.
 Scott, A et al. Front. NeuroInf., 2011
 Sikka, S. Resting-State, 2012
 Priebe, CE. arXiv:1112.5510
Preferred presentation format: