Mining Human Brain Data: Analysis, Classification and Visualization

Dartmouth College, Hanover NH

Investigators

Abstract

This multidisciplinary research effort develops methods for content based retrieval, analysis and visualization of probabilistic 3D spatial maps, in particular, the statistical parametric maps obtained from the fMRI (functional Magnetic Resonance Imaging) analysis of the brain. The approach is based on "activation signatures" which represent fMRI brain activations by size, shape, number (of foci), location, density, orientation, and other parameters. These signatures are used in providing automatic classification of activations into classes. Since fMRI brain imaging is a non-invasive technique it has attracted enormous interest in recent years because of its potential to solve fundamental problems explaining cognitive processes and to find relations between brain structure and function. The environment developed in this project can fully utilize fMRI's potential by enabling inter- and intra- studies and support large-scale mining of fMRI brain activations. The techniques can extract associations at multiple levels: among different activations for the same or different (groups of) individuals, between activations and tasks, between activations and ancillary data, and between tasks and ancillary data. Automatic characterization tools open the field to numerous new research opportunities, as they can offer common signature formats to do cross-modality brain data searches and correlation over multiple studies. In turn, the mined results can speed up discovery in the neuroscience field. The methods will be demonstrated on a prototype robust database of brain activations that includes scan data as well as biographical, clinical, historical and other subject-centered ancillary data. This database supports retrieval requests and subsequent analysis incorporating spatial statistics, classification, and visualization techniques and provides an evaluation of the system by the professional communities. The results of this work will be applicable to the analysis of other brain data such as brain lesions, Positron Emission Tomography (PET) scans, and MRI scans; thus helping to create tools for search across brain data formats. In addition, objects similar to brain activations appear in many other different domains, including cellular biology, tumor analysis, fluid dynamics, or financial records, therefore, the resulting techniques are expected to have broad impact. http://devlab.dartmouth.edu/